You type a question, read the answer, and move on with your day.

That’s how 9 out of 10 people use AI, and it’s why you think AI is overhyped.

The 1 that got the most out of AI is a pure gambler. And he’s winning the decade.

He thinks AI is a slot machine. And he’s right.

Once you accept that AI is a slot machine, you are at stage 5 (out of 5) of AI.

Because there are 5 stages people go through with AI, and they happen to map cleanly onto Kübler-Ross’s 5 stages of grief: denial, anger, bargaining, depression, acceptance. Guess what is also brutally changing our lives? AI.

Pick the line that sounded most like you this past week:

• “AI is overhyped.” → Stage 1

• “Look at this stupid answer ChatGPT gave me.” → Stage 2

• “I just need the perfect prompt.” → Stage 3

• “I’ve been at this for 6 hours and have nothing.” → Stage 4

• “I generated 100 answers, picked my favorite, done the job.” → Stage 5

I think most of you are stuck in stage 3 right now.

Now let’s get you up the ladder.

This newsletter grows from your shares. I want to help a million people use AI right. Sharing keeps your articles free & helps me stay laser focused on mastering AI for you.

Share

Stage 1. Denial

Denial comes in 2 flavors, and they look like opposites.

Flavor A: You tried ChatGPT once in 2023, it hallucinated a quote, and you’ve been telling that story to everyone ever since.

Flavor B: You paste outputs straight into client work because the answer sounded good. And ChatGPT said so. And ChatGPT is a super-intelligence, right? But then the client wasn’t so happy. So you stopped trusting the “super-intelligence”.

Both flavors come from the same blind spot.

AI is probabilistic. The answer you got in your chat right now was 1 of millions of options. It is both unique, and the average. Open a second chat with the exact same prompt, and you’ll get something different. That’s how the thing works.

How to get out of stage 1.

When you ask AI for something, ask for 3 versions in the same prompt.

Give me 3 different angles on [task] for [success criteria].

Make it your default way of working with AI = having options.

You’ll never go back.

Share

Stage 2. Anger

You got burned by AI. I’m guessing someone called you out once.

Now you screenshot bad AI outputs and send them to friends.

You probably posted a “look at this stupid answer ChatGPT gave me” thread on LinkedIn this year, or you bookmarked an article called “ChatGPT Got 60% of Medical Questions Wrong,” or the (fake) MTI brain scans showing how AI makes us dumb.

I get it. The first time AI hallucinates a source while sounding 100% confident, you feel lied to. Because in a way, you were.

But the anger is misplaced.

AI isn’t a truth machine. It’s an extremely competent junior intern: it follows your lead, very precisely. Maybe too much? It does not matter. You don’t yell at the junior, but the manager. You are the manager. Man up.

How to get out of stage 2.

When AI gives you something bad, just say it.

Yeah, I know. It sounds too simple. But most of my prompts are actually a follow-up that goes like this:

You missed the point. I meant to [call out the mistake].

I want to make sure you understood. So ask me clarifying questions.

Share

Stage 3. Bargaining

I bet this is where you are.

And where most of you will stay if you don’t read past this section.

*suspensful music because the content creator triggered you*

You think the perfect prompt is the answer. You’ve bought at least 1 prompt pack, you start every message with “You are a senior strategist with 20 years of experience...”, and you spend 20 minutes crafting a system prompt for a 30-second task. You believe prompt engineering is the key.

In a way, yes, good prompts are much better.

Like a great prompt raises your hit rate from maybe 5% to maybe 25%. The remaining 75% requires generating a lot more and throwing most of it away.

This is the plateau most of you hit. There is no “magic prompt that changed my life.”

Having the right setup for your prompt is good.

Understanding the principles of prompt engineering is awesome.

But my best prompts are actually quite small. I know the “AI magic” is in the conversation I’m going to have. And sometimes, it won’t work. That’s OK.

How to get out of stage 3.

I love this prompt for Claude Cowork (once you set up your folders correctly).

I explained how in this newsletter. I’m sorry. I promise it’s the only other newsletter you have to read. If you skipped it, you won’t have the same results.

Because I trained Claude Cowork on knowing who I am & my taste.

Once you do, and it’s saved in a folder, you can point to this folder when you start a new Cowork session, and simply prompt:

I want to [task] for [success criteria]. 
Start by reading my about-me folder and its files.
Then, use AskUserQuestion to complete your task with my answers.

Share

Stage 4. Depression

You’re generating a lot now. You also feel like you shouldn’t have to. You’re a little embarrassed about your credit spend, and your AI tab is getting pricey.

You sat down at 10 am, it’s 4 pm, you’ve generated 80 versions of one thing, and you have a quiet shame about it. You catch yourself thinking, “There has to be a smarter way,” so you bookmark another guide and wonder if you’re doing it wrong.

I’m sorry to break it down to you:

→ there is no smarter way, and the volume is the workflow.

To go from stage 4 to stage 5 is to accept “AI gambling” and regenerating like crazy. AI is expensive, I know. But that’s the only way to get something significant out of it.

The people shipping great AI work right now are generating way more than you think they are, and they’re doing it without apology.

I’ll show you a Midjourney example - an AI to generate images.

How come all of these images are fantastic?!

Because the guy who made it generated maybe 100, or probably 1,000 of them.

And the one you’re seeing is the final one. The best out of [x].

How to get out of stage 4.

AI subscriptions are like office supplies. It’s the price of work, of speed.

I don’t have a specific prompt to share.

But I do have a Claude skill called /hook to generate 40 Linkedin hooks before choosing one I like. And sometimes I don’t. So I ask for 40 more.

I have a consulting firm in NYC building Claude Skills for enterprises.

If you’re interested, send me an email at howtoai@rubenhassid.ai.

We can only take 1 company per week. Min. 50 people team.

Share

Stage 5. Acceptance (the playbook)

Stage 5 people generate in batches and expect to throw most of it away.

They’ve made peace with the math.

Here’s the workflow:

1. Write the task on one line.

“I need [task] for [success criteria]. But first, ask me questions.”

It might sound obvious, but you need to know what you want.

2. Pick a number.

How many versions do you need? Or options?

You usually need much more than you think.

3. Generate fast.

You want to make a lot of generations, and a lot of follow-up prompts calling out the mistakes. Until one sticks.

4. Download. Export. And do the final edits.

When I generate with Midjourney, I edit in Canva.

When I generate a spreadsheet with Claude, I edit it in Google Sheets.

When I write a newsletter with Claude Cowork, I finish the job in Substack.

AI is not to be lazier, but greater. You have time & options to find the best way to deliver something. Use this time accordingly.

Share

Either you don’t, or you overuse AI.

I only see two options.

1. You refuse AI categorically. And your product/service permits it. You are selling at an insane premium, and you have full confidence in your skill. No need for multiple versions, you already know the best way to do [x].

2. You use AI like a slot machine. You know you could never generate 100 variations of something so fast. And you also know the first shot of AI isn’t the right one. So you learn how to generate 100, fast, and pick the right one.

I respect People (1). And I am obviously People (2), most of the time.

But People (2) have some issues, and I want to be honest here:

• AI overload. We overwork ourselves. We manage multiple tabs at the same time. We do 4x more, 4x better, but we are also 4x more tired by 2 pm.

• AI addiction. Vibecoding is addictive. You try your luck with AI, and you build something you could have never made. And sometimes, you get a bit lost in the “passion work” instead of being actually productive.

I wish I could ask my 600,000+ readers how they deal with both: overload & addiction. You can leave a comment and share your own story:

Leave a comment

And for those who pay for my Substack, I made a post on Circle about it:

I will repeat it: you don’t have to pay for it. Ever.

Every one of my articles is free. And will remain free.

The Circle is to connect with me & my team & the other members:

Subscribe now

My mom always saw me too excited in front of MTV, watching video clips of The Arctic Monkeys, and kept repeating, “You know… they poop just like you.”

And my mom is right (moms are always right).

So I end up never glorifying people. They are just people.

But there is a small list of them I wish to meet very much:

• Ilya Sutskever, the previous OpenAI cofounder who made GPT possible.

• Steven Pressfield, the author who gave creators a battle plan against resistance and procrastination. You guys should read ‘The War of Art’.

• Andrej Karpathy, the previous OpenAI researcher who shaped early GPT models. Also, a fantastic educator. By the way, he just joined Anthropic (Claude).

• Rick Rubin, who strips creativity down to its raw essence.

• Robert Greene, who decoded power, influence, and human nature.

And then there is Demis Hassabis.

A Nobel Prize winner. Founder of Google DeepMind. A living legend of AI.

I just met Demis.

Shaking hands with the man building AGI.

So, I was at Google I/O on Wednesday, the 20th of May.

And you will quickly understand why it matters to you (too).

We talked for a couple of minutes about the future of work & AI.

If you have never heard of him, you need a quick recap.

• Age 4: started playing chess.

• Age 13: reached master level in chess. That’s very high.

• Age 17: helped create the video game Theme Park. A simulation.

• Early 20s: studied computer science at Cambridge + a PhD in neuroscience.

• Age 22: founded Elixir Studios, a video game company.

• Age 34: co-founded DeepMind in 2010.

• Age 37: DeepMind was acquired by Google in 2014 for 600 million pounds. It’s considered the best and cheapest acquisition of all time, in history.

• Age 39: led DeepMind when AlphaGo beat Lee Sedol at Go, the very best gamer of Go in the world, a game with about 10¹⁷⁰ possible board configurations, more than the number of atoms in the known universe.

• Age 44: DeepMind released AlphaFold, used for predicting protein structures.

• Age 48: won the 2024 Nobel Prize in Chemistry for AlphaFold.

And today, Demis is one of the handful of people building AGI (AGI = Artificial General Intelligence, when the machine becomes smarter than us, humans).

DeepMind has a simple mission: "solve intelligence, then use that to solve everything else." And Demis told me AGI - also called Singularity - is 4 years away from us.

Here’s what it means for your job over the next 4 years.

Share

PS: This newsletter grows from your shares. And I keep hitting 1,000+ shares! It’s my weekly north star. Sharing is free & helps me stay laser focused on mastering AI for you.

AGI is 4 years away.

Demis has been saying it for years. AGI by 2030. Maximum.

That’s what he believes. That’s what he’s actively building. And maximum means it could land earlier. Yes, I wrote 4 years. Not 40.

If he’s right, the work you do today looks almost nothing like the work you’ll be paid for in 2030. If he’s wrong, you have a few more years to get ready.

Either way, the move is the same.

And if you are wondering what I mean by “the future of work”, I want to take my dad as an example.

My dad’s marketing job (before spreadsheets).

My dad worked in marketing in the 1980s.

And I studied marketing from 2017 to 2022 (but dropped out, sorry, dad).

Needless to say, his marketing and my marketing have NOTHING in common.

My dad did hardcore marketing. Before spreadsheets. Before Google Forms. Before dashboards. Before any of the screens you’re using right now to do the same job. His job looked more like this:

• Walk into the street. Stop strangers. Ask them what they thought of a product.

• Write the answers in a notebook. Type them up later. Build the report.

• By typing them up, I don’t mean on a computer. I mean, on paper, of course.

• Draw mathematical functions by hand. BY HAND. With a pencil & compass.

• Walk it to the meeting. Present it to people, humans, in actual chairs.

Marketing was a physical job. A social job.

Then the screens arrived.

Spreadsheets. Google Forms. Looker. Tableau. Notion. Slack. Social media. Google Ads. The job wasn’t social. Got more analytical. Got faster. Got lonelier.

And now AI is taking that screen-work away.

The first draft. The summary. The market scan. The campaign brief. The deck outline. The Tuesday morning emails.

As per Demis own words, “Maybe we were not supposed to be watching a screen so much?” And I agreed.

That was - by far - the most touching moment of our discussion. I could see a very human researcher, hoping to build good for humanity, and wanting to solve the world’s best problem while making us more human.

I laughed - and you can see me laugh on the 30-second video - and said: “Somehow AI will push us to be more human than?” To which he replied, smiling, “Wouldn’t that be a nice future?”

So what’s left is what was always the real job. People.

The same place my dad started in the 1980’s.

Your job in 2030 will look more like my dad’s job in 1984 than it does today.

I asked Demis. And Demis agreed.

Share

I only pay my lawyer because I’d love to have lunch with him.

“Why am I paying my lawyer?”

He’s expensive. He drafts contracts. He flags risks. He sends me documents I sign too quickly.

If that was the whole job, I’d be using AI.

AI drafts contracts now. AI flags risks now. The output is fine. Sometimes good. Sometimes better than the junior associate he might hire next year. Especially with next year's levels of AI (imagine Claude-6, or ChatGPT-7).

But I still pay him. A lot. Why?

Because I want to have lunch with him.

I’m serious.

When he calls, I pick up. He has great insights, is nerdy about the media business, and makes my days easier. I want to team up with him. He’s fun to win with.

Contract is the floor, the minimum pricing. But our relationship is the ceiling.

And the ceiling is what costs me a lot of dollars per hour.

Now substitute your job:

• I pay my accountant because she remembers my mother’s name.

• I pay my designer because we have shipped hundreds of projects together, and she pushes back on me. Sometimes at 9 pm. Sometimes at 7 am. She cares.

• I pay my consultant because he said no to becoming my CMO. He told me, “You don’t need marketing help, stop searching”. My AI does not say ‘no’ to me.

• I pay my doctor because he remembers my back injury as if he can feel it.

The deliverable is the floor. You are the ceiling.

You, the human, the one with soft skills connecting with others.

→ If you stay on the surface level, the deliverable, AI will replace you in 4 years.

→ Become the ceiling, connect with others, be closer to the people you serve, and you get paid more than you’ve ever been paid in your life.

That’s the biggest takeaway from my conversation with Demis:

1. AI will absorb the need for most people to be in front of a computer.

2. Focusing on human interactions seems the most profitable way forward.

3. Taste & judgment are the only competitive advantages left. Sharp them.

But if you want a glimpse of the future, it’s called agent orchestration.

Share

The other path: agent orchestration.

I will be direct: most, if not everyone reading this shouldn’t be doing this.

Agent orchestration, or managing a swarm of AI agents to do the work.

In simpler terms, imagine an army of Claudes and ChatGPTs that can make decisions and do the work. 24 hours a day. 7 days a week.

Not today’s Claudes and ChatGPTs, but imagine the 2030 version of it.

I don’t know why people are obsessing over being Jarvis and controlling this swarm of “agents”. But I think - like Demis - that only a fraction of people will do it. And they won’t necessarily be the richest people around (or the happiest).

Just to be clear, what does it mean today to “orchestrate” agents?

You build a system of AIs that talk to each other. One agent does the research. Another writes the draft. Another sends the email. Another schedules the follow-up. You sit in the middle, like Iron Man with Jarvis in your ear.

Agent orchestration works. But it also takes weeks to set up, requires you to think like an engineer, and is mostly for the developer world.

That’s why I never write guides on “Agents” or “Automation”. It’s not there yet. But Demis told me, and knows, it will eventually become the norm.

But if you’re a knowledge worker in your 40s with a real job, a real calendar, and a real family, this is not where to spend your next 4 years. Developers will build the orchestration. You’ll use what they build, exactly like you use Excel today, without knowing how Excel was made.

Your premium move is somewhere else.

And Demis told me how to build your human premium today:

Share

Physical. Intelligence. Taste.

Physical labor is fixed. Intelligence is fixed.

Taste is the only human skill left.

Some say you can algorithmically produce taste.

I prefer to believe AI can hardly - very hardly - be tasteful.

AI can’t take responsibility for you. It’s not calling the shot.

The call to ship or kill. The call to hire her or pass. The call to bet on the deal or walk away. The call to send the email or sleep on it.

AI gives you 30 options in 30 seconds. Taste picks the one. Judgment knows when picking is wrong.

Let’s take our Nobel Prize winner, Demis.

Demis built DeepMind because he saw something in 2010 that the smartest people in the world didn’t see. He bet 16 years on it. They called him crazy:

But the bet paid: AlphaGo, AlphaFold, Nobel Prize.

That’s taste. That’s judgment.

You build them by doing the work. Years of bad calls. Awkward meetings. Failed launches. Clients who ghosted. Hires who didn’t work out. Projects that died.

AI has read about those things. You lived them.

That tiny little gap is your salary in 2030.

Pick up the phone. Show up in person. Remember the name. Push back. Take the long lunch. Make the hard call. Take responsibility when the work goes wrong.

Be more human to get paid a (massive) premium.

Do the things AI can’t do, because AI doesn’t have a body (yet????).

In 4 years, the average knowledge worker will be replaced by a subscription.

The premium knowledge worker is paid more than they have ever been paid in their life, because they’re the only thing in the room that’s still real.

Be that person we want to have lunch with.

Share

Am I paid by Google?

No.

They invited me to I/O, no strings attached.

Because I simply cannot take money from AI labs.

99% of my work is for free - shared to millions of people across social media.

And 99% of my work is about AI labs, like Google & Gemini, OpenAI & ChatGPT, or Anthropic & Claude. As I want millions of people to have a single source of truth when it comes to AI labs, I cannot take money from them.

So how do I keep giving for free?

1. By 2030, I believe 5 million humans will read my newsletter, for free. This newsletter will remain free to help you stay focused on how to use AI.

2. By 2030, I believe 50,000 of them (1%) will pay for my Substack. They join my Circle community, so AI isn’t such a lonely experience. I connect with them.

3. By 2030, I believe 500 of them (1% of 1%) will be my consulting clients - enterprises in the US - so I can help them accelerate their AI adoption.

In other words…

• You can do it by yourself for free.

• You can do it with you by paying for my Substack.

• Or I can do it for you, with my team, by messaging me on Linkedin.

I don’t care about Claude, ChatGPT, Grok, Gemini, or any other model.

I don’t pick sides. I’m not paid to make this newsletter.

I’m sharing, twice a week, how my work is transforming (very fast) with AI.

As I’m trying to keep up, I want you to keep up. So we move just as fast.

I want to be the greatest filter to the AI noise. And 600,000+ people read this twice a week to focus on the How. Some came because of my LinkedIn. But most readers subscribed because someone they trusted sent one of my articles to them.

If this article helped you, be that person for someone else (and share it):

Share

It does not cost you anything to share. And sharing is caring :)

If someone did send you this, thank them and subscribe for free here:

Subscribe now

How could I possibly know the best way to do it?

Because I tried every single AI to make spreadsheets.

1 - Copilot → absolutely terrible, couldn’t even make a spreadsheet.

2 - Copilot inside Excel → couldn’t make it work, after multiple trials.

3 - Claude → not as good as Claude Cowork.

4 - Claude inside Microsoft Excel → a good option. Not my favorite.

6 - Gemini (app and browser) → not as good as Claude Cowork or ChatGPT.

7 - Gemini inside Google Sheets → I wish, but I don’t have access to it (yet?).

8 - Grok - says “done” but you can’t open the excel file, you can’t even find it.

9 - ChatGPT → yes, but I prefer inside ChatGPT Google Sheets.

10 - ChatGPT inside Google Sheets → not to start, but to edit spreadsheets.

11 - Codex from OpenAI → it’s the best, but I was out of tokens within 2 prompts.

Only one is missing: Claude Cowork. Because it’s the best.

This newsletter is the full playbook of Cowork for Excel. How to set it up. How I use it every single week, deliver consulting work, and researching faster than I ever could alone. Also, where it falls short (I keep promising honesty).

Two things before we start:

1. Save this guide and spend 15 minutes this week to explore Cowork.

2. Send it to anyone who still hasn’t tried Cowork (or Claude).

Share

This newsletter grows from your shares. I want to help a million people use AI right. Sharing keeps your articles free & helps me stay laser focused on mastering AI for you.

Forget prompts.

Claude is no longer (all) about prompts, because of Cowork.

A quick Claude Cowork reminder for some of you + how to access it:

1. Go to claude.com/download. Download the app.

2. You need a Pro account ($20/month). It’s very much worth it.

3. Open the app. Click on the Cowork tab at the top between Chat & Code.

4. Select a folder from your computer. I wrote a lengthy guide on it here.

5. Make sure to always select “Opus 4.7” and “Adaptive thinking”.

Once you are good to go, it’s actually quite simple.

Share

How Cowork generates Excel.

I will explain both with a quick written list, and then with screenshots:

1. Go to the Cowork tab. Make sure Google Drive is connected (it will be easier later once we connect Claude to Google Sheets, I personally love it).

2. Use this prompt template:

Create an Excel spreadsheet from:
[DATA: file path, folder, or pasted data, best is to upload data].

### Purpose:
[Who uses this and what decision/task it supports — 1 sentence.]

### Sheets needed:
- "[Sheet name]": [columns, what each row represents, any calculations or formulas]
- "[Sheet name]": [e.g., summary with totals, pivot, charts — or remove if only one sheet]

### Formatting:
[Currency/date formats, conditional highlighting, frozen header row, totals row — pick what applies.]

Before building, list your top 10 assumptions so I can sanity-check them, then execute.

The magic part of the prompt is the “top ten assumptions to sanity check before execution”. I am the one controlling the AI, not the other way around.

Let’s now explain with screenshots from here so you can easily skim it.

Then I copied the prompt template I shared right before.

Here’s an example here for my own consulting firm:

Create an Excel spreadsheet: 12-month revenue forecast for an AI consulting firm, covering July 2026 – June 2027.

### Purpose:
Board-ready forecast to plan hiring and runway. Used live in board meetings — needs to look polished and let the board pressure-test every assumption.

**Business context:**
- 4 service lines with very different economics: (1) workshops, (2) full AI deployments, (3) AI sprints, (4) fractional Chief of AI retainers.
- Audience funnel (today, May 2026): Substack 600k growing +1m/year; social following 1m+ (800k from LinkedIn) growing +500k/year. Model how these feed inbound leads → discovery calls → proposals → won deals.
- Make your own assumptions for deal sizes, win rates, sales cycle length, delivery capacity, COGS, gross margin, and audience-to-lead conversion. Every assumption lives as a labeled, editable input on the Assumptions sheet — never hardcoded inside a downstream formula — so each one can be challenged and the model updates instantly.

### Sheets needed:
- “Assumptions”: all inputs grouped (Service lines, Audience & funnel, Costs, Headcount), with cell notes flagging which numbers are best guesses vs. anchored
- “Funnel”: monthly Substack + socials → leads → calls → proposals → won deals, split by service line
- “Revenue”: 12-month forecast by service line (bookings, recognized revenue, deferred), with monthly + cumulative totals
- “P&L”: revenue, COGS, gross margin, basic OpEx, contribution margin by service line
- “Dashboard”: single-screen summary with KPI tiles (ARR, total bookings, blended margin, headcount needed by quarter) and charts (revenue stack by service line, audience-to-revenue funnel)
- “Scenarios”: Base / Bull / Bear toggle that flexes 3–4 key drivers

### Formatting: 
USD currency (no decimals), percentages where relevant, frozen header rows, conditional formatting on margin %, clean palette suitable for board projection, named ranges for the major inputs so formulas read like English.

Before building, list your top 10 assumptions so I can sanity-check them, then execute.

It then took Claude Cowork quite some time to make the spreadsheets. About 7 minutes, I believe. But if it’s how long it takes to make the best spreadsheets possible with AI, not bad.

Now what’s next? You can continue following me inside Claude Cowork with prompts to modify the entire spreadsheets, or move to Google Sheets.

Why Google Sheets? I just prefer Google Sheets to Microsoft Excel. It’s easier to use & share with others. Just make sure your Claude is connected to your Drive:

Once your Claude Cowork is connected to your Drive, you can do this:

I just taught you how to start a spreadsheet from scratch with AI.

Now, let’s quickly discover how to edit one with AI. Surprise, I use ChatGPT!

Share

ChatGPT to edit spreadsheets.

So a quick recap:

1. We wanted to create a complex spreadsheet.

2. We downloaded Claude, paid for a plan, and went to the Cowork tab.

3. We already did our job to set up the folder & select the right 4.7 model.

4. And we simply copy & paste the prompt template I shared to make Excel.

5. We wrote follow-up prompts and made sure our Google Drive is connected.

6. Once we were proud of our spreadsheet, we clicked on “Google Drive”.

7. It opened a very solid spreadsheet, right inside Google Sheets (for free).

Technically, you probably know how to use spreadsheets and just want to manually edit from there. I won’t teach you this!

But if you want to edit with AI, there is a new ChatGPT extension in Google Sheets. And I actually love it! It’s very simple to use.

I obviously tested it vs. Claude Cowork. But Cowork is better to start from scratch.

It’s finally time to prompt it. Some examples:

Now the best use case: editing the spreadsheet.

I prompted ChatGPT:

In @Assumptions, change the “Bull” scenario to be much more optimistic. Make the right assumption on how much more would be still realistic.

This is my favorite setup.

I start with my Claude Cowork → then open Google Sheets + ChatGPT’s extension. But if your full-time job is making spreadsheets, I have another way.

Share

Another player: Shortcut.ai

I found another AI, Shortcut.ai, only made for spreadsheets.

I am not affiliated or paid by Shortcut.ai by the way. It’s an obvious reminder.

Shortcut wants to replace both Microsoft & Google. A bold move.

And it does not have the context my Claude has. But you could update it here:

You might want to ask me, “Ruben, what should I use?”

• I don’t make more than one spreadsheet a day. So Claude Cowork → to Google Sheets → to sometimes editing with ChatGPT extension is enough.

• But if your full-time job is making spreadsheets, it’s worth exploring other options like Shortcut.ai or ChatGPT Codex (also very good! but costly).

Share

Claude & Anthropic do not pay me.

So how do I keep writing everything for free?

1. Over 600,000 people read my newsletter for free. I don’t need 600,000 people to pay me. Most people will do everything by themselves, and I get them.

2. Over 3,500 people pay for my Substack. They join my Circle community, so AI isn’t such a lonely experience. I connect with them on Linkedin too!

3. My consulting firm, GPC, has already helped over 15 enterprises in the US to accelerate their AI adoption.

In other words…

• You can do it by yourself for free.

• You can do it with you by paying for my Substack.

• Or I can do it for you, with my team, by messaging me on Linkedin.

I don’t care about Claude, ChatGPT, Grok, Gemini, or any other model.

I don’t pick sides. I’m not paid to make this newsletter.

I’m sharing, twice a week, how my work is transforming (very fast) with AI.

As I’m trying to keep up, I want you to keep up. So we move just as fast.

I want to be the greatest filter to the AI noise. And 600,000+ people read this twice a week to focus on the How. Some came because of my LinkedIn. But most readers subscribed because someone they trusted sent one of my articles to them.

If this article helped you, be that person for someone else (and share it):

Share

It does not cost you anything to share. And sharing is caring :)

If someone did send you this, thank them and subscribe for free here:

Subscribe now

A subscription to an intelligence.

And the intelligent subscription is getting cheaper than your intelligence.

By cheaper, I don’t mean twice as cheap. I mean, 6,000x cheaper in 4 years:

A few years ago, if a company needed a decent memo, a market scan, a spreadsheet, a landing page, a sales email, a first draft, a summary, a slide outline, a contract review, or 20 ideas for a campaign, they needed a human.

Usually an expensive one.

Now they need a prompt. Maybe 30 seconds. Maybe $20/month.

And the output is… fine.

Sometimes good.

Sometimes better than the junior person they hired last year.

That’s why junior jobs are simply crashing:

And you can’t say it out loud.

→ The threat is not that AI becomes perfect.

→ The threat is that AI becomes cheap enough to replace good enough.

So the question is no longer: “Will AI replace me?”

Boring question. The better one:

What happens when the thing you used to sell becomes almost free?

That’s what we need to talk about.

Because if you understand this, you can stay ahead. If you don’t, you’ll keep asking for better prompts or tools while the market decides on a cheaper/faster option.

Two things before we start:

1. Don’t save this guide and forget it. Book 20 minutes in your week to act on it. Pick one section today and apply it to your actual work.

2. Share this article to your peers. I believe more people need to know. Do you?

Share

PS: This newsletter grows from your shares. And I keep hitting 1,000+ shares! It’s my weekly north star. Sharing is free & helps me stay laser focused on mastering AI for you.

The crashing cost of intelligence.

I was the smartest kid in school (up until highschool).

I was - and still am? - very proud of my intelligence.

I am not the tallest. I am not the strongest. But I definitely shine with intelligence.

So what happens when the thing that defines me the most gets so cheap I can no longer compete? I want to answer this daunting question today.

Most knowledge work was priced around 3 hidden costs:

1. The cost to know enough.

2. The cost to think through the work.

3. The cost to produce the final thing.

AI attacks all 3.

1. It knows enough to start.

2. It thinks enough to draft.

3. It produces enough to make you uncomfortable.

And yes, it still makes mistakes. But cheap tools don’t need to be perfect.

They need to be good enough for people to use them 100 times more.

When something useful gets cheaper, people use more of it.

• More drafts.

• More tests.

• More research.

• More prototypes.

• More analysis.

• More cold emails.

• More content.

• More dashboards.

• More legal first passes.

• More financial models.

• More everything.

It’s commonly known as the Jevon’s Paradox.

Jevon’s Paradox: Technological advancements increase the efficiency with which a resource is used, the total consumption of that resource often increases rather than decreases. This occurs because lower costs unlock new demand.

This is why I don’t think AI means “less work.”

I think it means more work gets attempted.

The lazy prediction is: AI replaces people. The better prediction is: AI increases the amount of work humans are expected to manage.

And that creates a new problem → intelligence is becoming a commodity.

You won’t lose because AI can think.

You’ll lose because everyone around you can suddenly think 100 times more often, and you’re still working like in the past, when thinking was expensive.

Share

Intelligence used to be a bottleneck.

Think about how work used to happen.

1. Someone had an idea.

2. Then they waited.

3. Waited for the analyst.

4. Waited for the designer.

5. Waited for the writer.

6. Waited for the intern to summarize the report.

7. Waited for the consultant to make the deck.

8. Waited for the developer to build the prototype.

9. Waited for the manager to review the memo.

10. Every project had bottlenecks.

11. And many of those bottlenecks were just expensive intelligence.

12. Someone had to read.

13. Someone had to think.

14. Someone had to write.

15. Someone had to format.

16. Someone had to check.

Now the first pass is (almost) free.

→ That doesn’t mean the final work is free.

→ It means the first (serious) attempt is free.

Huge difference.

Because once first attempts are cheap, the number of attempts explodes.

Before AI, you might test 1 headline.

→ Now you test 50.

Before AI, you might read 3 competitor pages.

→ Now you scan 300.

Before AI, you might make 1 client deck.

→ Now you make 4 versions for 4 different buyers.

Before AI, you might ask one analyst for one answer.

→ Now you ask 10 questions you never would have asked because it wasn’t worth anyone’s time.

That’s why “AI is cheap” really means “impossible volume is now possible”.

And if the volume changes, the job changes.

Share

Average dies first.

AI is very good at average.

• Average memo.

• Average sales email.

• Average research summary.

• Average content calendar.

• Average Python script.

• Average slide outline.

• Average customer support answer.

• Average job description.

• Average legal summary.

• Average market analysis.

• Average “strategy.”

I’m using the word average on purpose.

Average used to be enough.

Average looked professional because it took time.

Now, average looks suspicious because it took 12 seconds.

Did you just use ChatGPT to do this?

The worst sentence you could receive, from anyone. Especially a client.

This single mentality change is the beginning of the collapse of (some, or most?) knowledge workers. They used to have a moat made of effort.

1. You had to sit there.

2. You had to type.

3. You had to know the format.

4. You had to search.

5. You had to assemble the pieces.

AI removed pretty much everything. And is doing it even faster. Even cheaper.

So now the market looks at average and thinks, “Why am I paying a human for this?”

Painful. But also useful.

It tells you exactly what to do next:

Stop trying to make average work. Move above it.

Share

Judgment is the new work.

When intelligence gets cheap, judgment gets expensive.

Everyone can generate options → Few people can pick.

Everyone can draft → Few people can decide what should exist.

Everyone can summarize → Few people can tell what matters.

Everyone can make a slide → Few people can make the one that closes.

The valuable person is no longer the person who can produce the first version.

The valuable person is the one who can look at 20 versions and say:

• “This one.”

• “This angle.”

• “This risk.”

• “This is wrong.”

• “This will fail with our buyer.”

• “This sounds good, but solves nothing.”

• “This should never be sent.”

That’s taste. And taste is trained by doing the work.

Thousands of reps. Bad calls. Awkward meetings. Failed launches. Clients who didn’t buy. Readers who didn’t care (hi! if it’s you). Bosses who killed the project. Users who clicked the wrong button.

AI has read about those things. You lived them.

It’s an important gap, that will get more and more expensive.

You must find a way to extract your taste and give it to AI.

Because if your taste stays trapped in your head, AI can’t work with it.

You need to turn your expertise into something AI can work with:

Share

Your expertise needs to become files.

If you want to stay ahead, start turning your expertise into reusable context.

It goes beyond simply not repeating “I work as a [job] at [company]”.

It’s giving AI as much context as you can. What you love, hate, the secrets of your best work. Your standards. Your taste. Your scars. Your audience. Your rules. Your mistakes. Your constraints. The weird little things only you know.

Put them in markdown files (a text file AI loves) → then give them to AI before important tasks → tou’re feeding AI the part it can’t scrape from the internet → that’s how you stop getting the average answer.

A prompt is a one-time request. But a file (inside a folder) is reusable memory.

So I keep coming back to this, because it works.

Here’s the folder I built for my Claude Cowork. But it works with other AI by just uploading all the files into a “Project”.

I wrote an extensive, free guide on how to set up Claude Cowork here.

I know it’s annoying to read a newsletter that asks you to read another newsletter. But that would make each and everyone of them horribly long. I’m sure you understand.

ABOUT-ME/
- about-me.md
- anti-ai-writing-style.md
- my-company.md

TEMPLATES/
- examples-i-love.md
- examples-i-hate.md
- [whatever example of work you love here]
- [whatever example of work you love here]
- [whatever example of work you love here]

OUTPUTS/
- [whatever AI produces, it's here]
- [whatever AI produces, it's here]
- [whatever AI produces, it's here]

Keep it small. Please don’t dump 400 PDFs into a folder and call it context.

It looks like this inside Claude:

Now that your folder is ready, time to prompt it:

Share

A good prompt is actually quite simple…

… if you have the right folder and file system.

Just open Claude, ChatGPT, Gemini, whatever you use.

Paste this:

I want to [task] for [success criteria]. 
Start by reading my about-me folder and its files.
Then, ask me questions to complete your task with my answers.

This will feel slow. Good. It takes time to transfer taste.

Most people want AI to skip thinking. You want AI to extract your thinking.

Once you do it on a large scale, you start understanding where most of your value is: it’s not the manual work, but your unique way of thinking.

So stop selling hours. Start selling taste.

Share

Stop selling hours.

“I spent 6 hours on this.” “I worked all weekend.” “This took me 3 days.”

Nobody cares.

I say this with love because I’ve done it too.

Time spent used to signal quality because production was expensive.

Now production is cheap. So time spent is a weak argument.

The better argument is:

• “I understood the real problem.”

• “I knew what to ignore.”

• “I found the missing risk.”

• “I made the decision easier.”

• “I shipped the version that worked.”

• “I saved the team from doing the wrong thing beautifully.”

That is what stays expensive.

Outcomes. Taste. Trust. Responsibility.

Move from answer person to system person.

The answer person gets replaced first. Someone asks a question.

→ They answer.

Someone needs a draft.

→ They draft.

Someone needs research.

→ They research.

That work used to be useful. Now AI does much of it.

The system person asks a better question: “Why do we keep needing this?”

Then they build the repeatable way to handle it.

• A template.

• A checklist.

• A context file.

• A project folder.

• A recurring review.

• A workflow.

• A trained AI setup for the team.

• A way for other people to get 80% of the answer without asking them.

That’s how you stay ahead.

You stop being the bottleneck. You remove bottlenecks.

Here’s the prompt:

Here are 5 tasks people keep asking me to do:

[PASTE TASKS]

That's exactly how I do it, step by step, with each tools explained:

[PASTE YOUR EXACT STEPS]

For each task:
1. Identify the repeated pattern
2. Tell me what context files are needed 
3. Turn it into a reusable checklist
4. Draft the first version of the workflow
5. Tell me what should stay human

AI builds systems. And I’m not talking about the monster-automation sending emails on your behalf to your loved ones. But simple recurring tasks.

The key play is to work with AI on what you can judge best.

Share

Use AI where you already have taste.

People use AI backward.

They go: “I’m bad at this. AI will do it.”

Dangerous.

If you’re bad at the thing, you can’t judge the thing.

You’ll approve clean garbage. Use AI where you already have taste.

If you’re great at sales, use AI to draft 20 sales angles.

→ You’ll know which 19 are weak.

If you’re great at finance, use AI to test your model.

→ You’ll catch the lazy assumption.

If you’re great at writing, use AI to produce raw material.

→ You’ll cut the fake voice.

If you’re great at operations, use AI to map workflows.

→ You’ll see the missing handoff.

Your skill becomes the steering wheel.

Because yes, humans still win.

You stay ahead by getting good at the things cheap intelligence doesn’t automatically solve:

1. Knowing what should exist.

2. Knowing what should be ignored.

3. Knowing who needs to trust it.

4. Knowing what will fail in the real world.

5. Knowing when the answer sounds right but is wrong.

6. Knowing how to make the work fit your exact context.

7. Taking responsibility when the output matters.

That last one is the big one.

I pay my lawyers to take responsibility. Be like my lawyer.

Share

AI is a lonely experience.

You’re facing your computer, by yourself, at work.

You can’t really brag about using AI. But you can’t really skip it either.

So you quickly open it and try to do your best work, by yourself.

You can’t ask your coworkers how to do better. It’s touchy. It’s literally your job. And the entire company keeps pushing for “AI champions”.

AI is such a lonely experience.

That’s why I started this newsletter, ‘How to AI’.

I was trying to make sense of LLMs (like GPT), but only devs had a community.

I do believe anyone reading this newsletter - for free - can make the most of it. Read, twice a week, a completely free guide on how to use AI. Book a 20-minute meeting with yourself every week if needed.

This is the “I do it myself” option.

Now, if you want the “Do it with us” option, join my Circle.

You’d have to pay for my Substack (it’s $200/year), and you will get:

1 - An invite to my private Circle channel (this is the best value for money).

2 - $219 worth of AI tools I partner with (like Gamma, Wispr Flow, Granola…).

Subscribe to my Substack to join my Circle here:

Subscribe now

P.S. There is also the option of “I do it for you.”

I consult enterprises in the US to make them adopt AI (Claude) faster.

We run workshops for teams of at least 50 people.

We deploy engineers where automation is needed.

We make sure AI works, and keeps working for people (your best asset).

I have already taken care of 14 businesses since January, but we are not live yet!

If you want to make sure to have a spot before I go live, shoot me a DM here:

My team will make sure you qualify first!

Until then, I am rooting for you, the human, to master AI before it masters you.

Humanly yours,

Ruben.

This is how most people (you?) outsource to AI:

Just like you outsourced your orientation to Google Maps, and you got dumb.

I am so bad at orientation myself.

Like, I am from Paris. I have lived most of my adult life there. But I can’t walk around without a GPS. I can’t help it: I put my destination, I follow the little blue dot, and I don’t even look up anymore.

It’s actually so bad I believe if you teleport me to Paris in the 80’s, without a phone, I’d be as lost as a tourist.

That was me & orientation. I could also mention…

1. Phone contacts → I know mine, and my mom’s. That’s it.

2. Spelling mistakes → Writing without an autocorrector isn’t fun, at all.

3. Mental math → Quick maths: 180 x 45? Yep. You’re opening your calculator.

4. Attention span → I remember how slow 3G internet was. If I were given 3G today instead of my regular 5G, I would go crazy. That’s ridiculous.

5. Photography → I expect everything to be on my phone. It’s stored, forever. So I stopped checking my old photos. There are too many. Lost the charm.

6. Calendar → I remember my family birthdays, a couple of friends. But if anything isn’t on my calendar, my brain is now really good at skipping it.

7. Music → If it’s not on Spotify or recommended by the algorithm, I don’t dig for novel songs anymore. But that used to be my job as a label owner.

You know where I am going with this. Today, there is AI.

Let’s say you face a new problem. You open AI. You can outsource your thinking to AI. But you can’t outsource your understanding.

You don’t want AI to take over the same way Google Maps made me orientation-dumb. And you’ve been there before, the AI “making you dumb”.

• You fix a slide deck because the AI version was off.

• This client’s email you rewrote 3 times because Claude’s draft sounded like a stranger. You imagine the client calling you out for using AI… you can’t.

• The spreadsheets you quickly built for your team, but you now have to cross-check every formula on every tab. Manually.

You’re using AI because it saves time → and you end up wasting a lot of it.

But there’s a way to use AI without outsourcing your brain. It costs you a couple of extra minutes of thinking on the front. It saves you hours on the back. And the work that comes out the other side is yours, not Claude’s.

I will share my own real-life scenario, 5 steps, as a detailed use case. Then, 3 more examples at the end of this newsletter, so you can see the same method run across very different work.

This is the guide I wish someone gave me before I wasted months trying to outsource everything to AI.

1. Save this guide and spend 30 minutes this weekend to ask AI better.

2. Send it to anyone asking you, “AI makes me dumber, how can I do it better?”.

Share

This newsletter grows from your shares. I want to help a million people use AI right. Sharing keeps your articles free & helps me stay laser focused on mastering AI for you.

I - How I outsource work (not thinking) to AI.

Last month, I got on a video call with the Chief of Staff.

240-person company. And 25 minutes in, she said: “I need to roll out Claude across the whole org. Help me sell you internally.”

By Friday, she wanted 3 things from me, all written in her voice, that she could forward to her CEO and exec team to get the rollout approved:

1. A 90-day Claude adoption roadmap.

2. A 1-page pricing breakdown (her side, my side, Claude side).

3. An email she could send to her exec team to kick it off.

Not to mention her credibility partially on the line for the recommendation.

The default version of me opens Claude and types: “Write a 90-day AI adoption roadmap for a 240-person company.” Maybe write a couple of lines about the company as “context” to feel good. Hits send. Copies the answer into a Word doc.

And it’s boring. Generic. Her CEO would smell it. Should I start again?

But here’s what the new version of me does:

Step 1. Give Claude more context than you think.

We must answer 4 questions before doing the work.

1. Who’s the audience?

2. What’s the one thing they should walk away with?

3. What format am I copying? (Paste a real example. Not “don’t sound like…”)

4. What are the stakes?

For my example:

1. Audience. 3 layers.

   • The Chief of Staff. 4 years in role, came from PE. Hates being sold to. Minute 12 of the call: “If this rollout fails, my CEO wants to contact Anthropic’s team.” Everything I send her has to land as risk reduction.

   • Her CEO. Reads the first paragraph and the dollar figure. That’s it (the Chief of Staff said so on the call). Phase 1 of the roadmap and the pricing one-pager have to land in 90 seconds of skim.

   • Her 8-person exec team. 2 former MBB consultants (they’ll dissect the methodology). 3 operators (allergic to slideware). 3 technical (will ask about SOC 2 and data residency on slide 1). I write the email for the most skeptical of the 8: the CTO who shipped 2 failed AI pilots in 2024.

2. Walkaway.

   • From the CEO: a 1-line Slack to the Chief of Staff that says “approved, send me the contract.”

   • From the CTO: a private DM to the Chief of Staff: “this one’s different. I’m in.”

   • It sounds cliché, but Claude is extremely good at understanding goals.

3. Format.

   • Roadmap = the 90-day plan I built for a similar engagement that closed 6 weeks ago. Phase 1 has to be 1 page. Phase 2-3 can be denser.

   • Pricing one-pager = her finance team’s vendor procurement template (she dropped it in our chat during the call).

   • Email = 180 words max. No bullets. 2-word greeting max. (Her last 5 internal emails were all under 200 words. No warm-up.)

4. Stakes. The Chief of Staff mentioned a potential co-hosting event, to invite their clients. And most of their clients are potential clients for us.

That’s about 12 minutes of thinking.

Some answers came from memory. Most from a Claude chat with connectors to my notetaker, my email, and my team’s channel. We must capture everything.

But where? Claude has so many different faces:

• Claude chat (browser).

• Claude chat (app).

• Claude chat Projects (app).

• Claude Cowork (app).

• Claude Cowork Projects (app).

• Claude Code (app).

Let me walk you through the best way to do it.

Now we need to extract my meeting notes, my emails & the Slack thread conversations I had about this specific client.

Here’s my prompt:

Use my Granola, Gmail and Slack Claude Connectors.

Extract all of the data from my call and emails threads and conversation with [my partners] regarding [client].

I want to answer:

1. What's the need?
2. What's the audience?
3. What's at stake?
4. What's the format they want?
5. What's the blocker?
6. What's the walkaway we want?

And it looks like this:

Let’s keep prompting in the same Claude Projects.

We can now start explaining what we need from Claude.

But we’re still in research mode:

I want [deliverables] for [stakeholder].

But I need to run a deeper analysis to catch what our conversations and meetings didn't catch. For example: 

1. I need [company] last annual report and 3 recent press releases. Is it still relevant? Does it help us?
2. What else could you find from deep searching the web on [stakeholder] and [company]?

Make sure you turn on Claude Research before sending the prompt:

That’s it for the prep work.

And you’ve now done more prep work than 99% of the people opening Claude. And you still haven’t asked for a single sentence of the draft.

Step 2. Ask for the thing.

This is actually the simplest step.

Here I have 3 deliverables. So I will ask 3 times:

Now that you have all of the necessary context, I need to send to [stakeholder] a [deliverable] for [success criteria] in [format].

I prefer asking 3 times (so 3 different prompts) instead of 3 asks in one prompt.

You simply get more reasoning, and a better answer every single time.

PS: If you didn’t turn on Opus 4.7 + Adaptive Thinking, it’s time to do it:

Step 3: Treat the first answer like a draft from a junior intern.

Claude’s first answer is rarely the right answer. Useful as a draft. Dangerous as a deliverable. Don’t accept it. Push back before her CEO does.

3 prompts I love to use.

Move 1. Roast it. Paste the output back in:

What's weak about this roadmap? Where would the CEO push back?
Give me 5 honest critiques as if you were the CEO trying to reject this.

Then make Claude fix what Claude just wrote.

Move 2. Reverse the brief. Ask Claude:

What did I forget to tell you that would have made this stronger?

This is where I find out I didn’t include her industry’s compliance constraints. Step 2 didn’t quite hold. Back to the folder. Add the SOC 2 documents. Restart.

Move 3. Compare to gold. Paste in the roadmap I built for the client that closed last quarter:

Why is your roadmap worse than this one? Rewrite to match the structure and confidence.

For this engagement, I run all 3 moves on the roadmap. Move 1 on the pricing one-pager. All 3 on the email.

Step 4. Follow up much more than you think you should.

1-turn AI chats won’t get you far.

The good stuff is in the follow-ups. Turn 12. Turn 27. Turn 41.

Two tricks make this fast/cheap enough.

Wispr Flow. It’s a dictation tool. You talk, Wispr Flow types.

I stopped typing follow-ups 2 years ago. I talk to Claude. 200 words per minute instead of 60. Each follow-up costs less effort. So I follow up much more.

I wrote a newsletter on how to set up Wispr Flow.

AskUserQuestion (inside Cowork). Ask Claude to use AskUserQuestion to ask you questions. It’s my favorite Claude feature. Here’s an example:

Instead of me prompting Claude, Claude asks me 3-4 questions and builds from my clicks.

For this example, I run roughly 40 follow-ups across the 3 deliverables.

The non-obvious ones:

• “Argue against this roadmap as the CFO who hates new vendor spend.” I take the strongest counter and bake it into Phase 1.

• “What would the IT director secretly fear about this rollout?” I quietly remove the part she’d fear.

• “Rewrite this email line by line as if she wrote it. Not me.” This one alone takes 10 follow-ups.

• “If this rollout fails by month 3, what’s the most likely reason? Now bake the prevention into the roadmap.”

If you finished an AI task in 2 turns, you stopped too early.

Step 5. Read the whole thing one more time before you ship.

Even after 47 follow-ups, the final output has tells. AI rhythm. A banned word. A bullet that doesn’t earn its place. A claim you can’t back up.

So this is what I did:

• I read the email out loud, twice.

• I read the pricing one-pager as her CFO: are the numbers defensible? Are the assumptions clear?

• About 25% of the original draft gets cut.

I usually delete stuff instead of writing new ones. AI writes too much, details too much. And I prefer that (so I can clean instead of adding new things).

On Friday morning, I send the Chief of Staff the 3 needed documents. About 2 hours of total focused work across a few days.

Now ask yourself the obvious question:

II - “If it’s that long, why even use Claude?”

Because you can’t hold 30,000 words of client context in your head.

Claude can.

You can’t run 3 versions of a roadmap through 3 different stakeholder lenses in parallel.

Claude can.

Set up your Cowork folder once and the next engagement for the same client is half the work. Use Wispr Flow and AskUserQuestion right, and 40 follow-ups feel like 5. Save your folders and prompts as templates, and the third engagement of the year takes 20% of the time of the first.

Claude takes each of the 5 steps from 10 hours down to 20 minutes.

Up to you to use the free time to do better work now.

AI was never meant to be lazier, or faster. But making you more capable.

And most people are not more capable with AI.

This is how most people outsource to AI:

I shared one example of how to use AI properly. Let me share three more:

Share

This newsletter grows from your shares. I want to help a million people use AI right. Sharing keeps your articles free & helps me stay laser focused on mastering AI for you.

III - 3 more examples of outsourcing to AI.

You’ve seen the method on a heavy consulting engagement. Here’s the same 5 steps run across 3 very different jobs. But I won’t explain so much this time.

Use case 1. LinkedIn post for a life coach.

The setup. A life coach wants to teach corporate teams to think like elite athletes. She needs 1 LinkedIn post a week. And she needs to use AI correctly.

Step 1. Audience = burnt-out middle managers in tech. Walkaway = 1 athlete habit they can copy tomorrow morning. Format = her viral post from June (paste it). Stakes = newsletter signups.

Step 2. Context dropped into Claude: her last 30 LinkedIn posts (so it learns her voice) from her Google Drive connected to Claude, 3 long-form interviews with the athlete she’s referencing (so the quotes are real, not hallucinated), the corporate burnout stat she always cites, and her tone.

Step 3. Roast move only. 1 paste-back: “Rewrite this with the punch of her best post.” Done in 2 minutes.

Step 4. 10 follow-ups. Mostly cutting and sharpening the hook. Wispr Flow makes it 4 minutes.

Step 5. She reads the post aloud. If she stumbles on a line, we rewrite the line. 30 minutes total from idea to scheduled.

Ship: 1 post in her voice, scheduled for 9am Monday. It looks like this:

Use case 2. A 30-page market report into a deck.

A logistics partner asked me whether they should expand into the peptide market. Deliverable: a 30-page report and a board-ready deck.

Step 1. Audience = their board (8 people, 6 of whom are 60+). Walkaway = a clear yes/no with the 3 strongest reasons either way. Format = a sample McKinsey market-entry report (paste it). Stakes = the engagement plus a retainer if they say yes.

Step 2. Context: their last 2 annual reports, the FDA’s recent peptide guidance, 4 peptide-market analyst reports, financials and ops summaries on 6 competing logistics players, transcripts from 3 peptide companies they could service. About 180,000 words into a Cowork Project.

Step 3. I like to sometimes roast Claude by using ChatGPT. I will go and use ChatGPT’s Deep Research (its own Research) mode and add it to my Claude Cowork Project. I will then ask, “What did you forget that ChatGPT didn’t miss?”

Step 4. Dozens of follow-ups across multiple Claude chats (inside the same Project). Each market segment gets its own thread.

Step 5. I read the full report in 1 sitting on paper. Spot-check 10 random data points (do the sources hold?).

Once done, I make sure my Gamma connector (to make slides) is active.

And I simply prompt “Use the Gamma connector to make the necessary slide”.

Ship: 30-page report and 22-slide Gamma deck.

Use case 3. Spreadsheet sent to a CTO.

A CTO asked me whether her 50-person team should invest in building their own open-source LLM stack (Llama or Qwen on H200 GPUs) or just keep paying Anthropic for Claude API. He needs a financial model showing token economics over 6 months.

Step 1. Audience = his CFO and the board. Walkaway = the break-even point in months, and the answer. Format = a 4-tab Excel model with sensitivity analysis (paste a model that closed a board vote last year). Stakes = $3M+ in CAPEX or 6 years of API spend, depending on the call.

Step 2. Context: current Claude API tiered pricing, open-source model benchmarks (Llama, DeepSeek, Qwen), H100/H200 GPU CAPEX vs. AWS rental rates, salaries for the 3-person ML infra team they’d need to hire, 12 months of her team’s actual token consumption logs.

Step 3. Move 1 only. “What did I forget in this model? Where would my CFO call BS on the assumptions?” Claude flags 3 missing line items (fine-tuning costs, model retraining, evaluation infra). I add them.

Step 4. 10 follow-ups. Most of them iterate on scenarios: base case, best case, worst case. Plus 3 stress tests. What if Claude pricing drops 50%? What if open-source quality matches Claude next year? What if GPU prices double?

Step 5. I rebuild the spreadsheet manually for 1 month of usage.

Ship: 4-tab Excel model and a 1-page executive summary.

Share

I just opened a community, and 500 people joined.

But I don’t want (all of) you to join.

I gatekeep this community to people paying for my newsletter.

I know you are one of my 600,000 free subscribers. So why is there a paid version?

Because it protects us from bots, spammers & abuses.

I will repeat it: you don’t have to pay for my newsletter. Ever.

Every one of my articles is free. And will remain free.

But if you do want to pay (it’s $200/year), this is what you will get:

1 - An invite to my private Circle channel.

2 - $219 worth of AI tools I partner with (like Gamma, Wispr Flow).

3 - Live Q&A and access to our workshops (mostly about Claude Cowork).

If you want to upgrade today, go here:

Subscribe now

Because my Claude is connected to everything else: my Gmail (email), my Slack (team channels), my Granola (meeting notes). It’s called Claude Connectors.

Claude has over 200 connectors. But you’re probably using none of them, and this newsletter is here to help you get started.

Two things before we start:

1. Save this guide for later. Book a 15 min meeting this week to act on it.

2. Share it with your team who wants to get more from Claude.

Share

PS: This newsletter mostly grows from your shares. 1,000+ shares per article is my weekly north star. It keeps the article free & helps me stay laser focused on mastering AI for you.

I - How to connect your app to Claude.

A Connector is a 1-click bridge between Claude and an app you already pay for.

Here’s how to connect your app to Claude:

1. Open Claude. Go to the plus “+” inside the chatbox.

2. Find “Connectors” and click on “Add connector”. Add one of them.

3. Now open a new chat. Click “+” in the chat bar > Connectors > make sure your app is toggled on for that conversation. It should be automatic.

What changes when you turn one on?

Let’s see some examples:

• Gmail on: ask “what’s pending in my Gmail inbox from this week?” and Claude answers from your actual Gmail. 0 copy-pastes.

• Granola on: ask “I need a summary of all of last week’s meeting transcripts,” and Claude reads them all. 0 dashboards opened.

• Slack on: ask “extract this entire thread and give me the action items,” and Claude pulls the entire context. 0 tabs visited.

PS: 3,500 people pay for my Substack to get access to me in a Slack channel. But the Slack got messy. So I am migrating to Circle instead. If you are a paid subscriber to this newsletter, hit me up by DM! I’ll send an invite to my Circle.

Share

II - My 9 favorite Claude connectors.

Now here are my 9 favorite Claude Connectors I (actually) use every week.

1. Granola. For my meeting transcripts.

Granola is my notetaker. I think it’s the simplest/most complete one out there.

When I join a meeting (on Zoom, Teams, Meet), Granola automatically captures the entire transcript and stores it on my computer inside its app.

This is a typical prompt I use inside Claude, with Granola:

Pull my last meeting from Granola. Summarize in 5 bullets.
Then list every action item assigned to me, with deadlines.
Then flag the 2-5 blockers I might have to complete some items.

Pro tip: Combine Granola’s context with everything else. For example, when I’m negotiating a contract, I ask Claude to read Granola+Slack+Gmail. Combine the connectors together!

2. HubSpot/Salesforce. For lead generation.

Most people use HubSpot or Salesforce to manage their leads (in a CRM).

Prompt I use inside Claude:

Read my HubSpot pipeline.
List every deal in stage "Proposal Sent" that hasn't moved in 7+ days.
For each: write a follow-up email I can send today.
Tone: direct, French-blunt-warm. No "just checking in."

Pro tip: Stop using HubSpot/Salesforce only as a CRM. Use it as a research input. Ask Claude, “Based on my last 5 closed-won deals, what’s the common thread?”

4. Notion. For my team’s knowledge base.

I use Notion for our content calendars and knowledge base. Like, for example, my social media posts, all of our standard procedures on how we do things…

Prompt I use with Notion:

Go to Notion, NEW Post Calendar. Extract the content from May 12, 2026 (Design brief, Caption). Create 2 new ones using the same ideas and techniques.

Worth noticing, I am a small company, but most big ones use Microsoft or Atlassian with Jira, Confluence... Needless to say, Claude can connect to all of them. But I haven’t tried. So go & do it and let me know if it’s good :)

5. Microsoft 365. Because you probably have to use it.

One toggle. 4 apps: Teams, Outlook, Calendar, OneDrive.

I consult SME businesses in the US, generating at least $100 million per year. I think most, if not all, of them were using Microsoft. You’re probably one of them.

This is a typical prompt to call the entire Microsoft suite:

Read my Outlook + Calendar for next Monday.
List every meeting where I have no prep notes yet.
For each: pull the relevant Teams thread or OneDrive doc.
Output a 1-page brief per meeting.

Pro tip: It works even better when the entire company adopts it. Because then people adopt new ways of collaborating/communicating. They know AI can use it, so they start being clearer everywhere. This is the new way of working.

6. Slack. Because I don’t use Microsoft Teams.

I use Slack, not Teams, to talk with my teams.

A typical prompt I use :

Read #tathastu, #team-lead, #gpc-consulting, #circle-studio from the last 48 hours. Surface the 3 things I need to react to. Skip jokes, GIFs, standup updates.

7. Google Drive. Because I know we all use Google Drive.

I use Google Docs & Google Sheets heavily. So I have to connect Claude to my Drive. But the way I use it with Claude here is a little different.

1. I go to Claude Cowork. And I explain the spreadsheets I want.

2. Cowork builds it. And I keep refining it with follow-up prompts.

3. Then eventually, I click on the “Google Drive” button to open a new tab.

4. Now all of Claude’s work is on my Google Drive (= Google Sheets here).

8. Gamma. To make on-brand slides in seconds.

I use Gamma for all of my slides. It’s the fastest/cheapest/easiest. Hard to beat.

A typical Gamma prompt:

Go to Gamma. Create a [NUMBER]-slide deck on [TOPIC] for [AUDIENCE].
Goal: [What you want them to do or understand].
Cover: [List 5-8 key points or sections].
Tone: [confident / data-driven / concise].
Make it visual, clean, and one idea per slide.

9. Gmail. The most obvious win.

I save the most known for last.

What it replaces: searching your inbox, finding the wrong email, searching again.

It also works for Outlook emails, obviously. A prompt you’d like to try:

Read my Gmail from this week.
Group every email into 3 buckets:
1. Needs my reply today.
2. Can wait until Friday.
3. Newsletter / FYI / can ignore.
Output: 3 lists. No commentary.

Non-obvious tip: ask Claude to draft the replies as drafts in Gmail, not in the chat. Just add the prompt: “Save the 5 replies as drafts in my Gmail.”

If you only turn on 3 connectors today, pick Granola, Gmail/Outlook, and Slack/Teams (depending on what you use internally). This trio compounds.

Share

III. The entire list of Claude Connectors.

The directory is 200+ Connectors and growing.

If you want to find all of them directly inside your Claude, do this:

Share

IV. When NOT to use a Connector.

3 rules.

1. Don’t leave them all on by default.

Connectors burn tokens before you type. Every Connector loaded is context Claude has to scan.

Turn them on per chat. Off per chat. Treat the toggle bar like a kitchen tap.

If you need more tips on avoiding wasting Claude’s tokens, click here.

2. Don’t use Connectors for creative work.

When you’re drafting a LinkedIn post or writing a script, you want a clean room.

Slack noise, Gmail noise, Notion noise pollutes the output. My default writing chat: every Connector off. Always.

3. Don’t grant write access until you trust the read flow.

Read-only first. Watch how Claude uses it for a week.

Then grant write. Most data accidents happen on day one when someone enables a Connector with full write access and asks Claude to “clean up my Drive.”

Share

V. Build your own connector (for devs).

This section is very technical. 99% of you don’t need it. So skip it.

A Connector is just an MCP server with a public URL.

You won’t create one. But your tech team could!

Or ask me to do it for you :) I have a consulting firm in NYC.

You can ship one in maybe 3-4 hours if you’ve never touched code. 15 minutes if you have. You’ll need:

• A Claude Pro, Team, or Enterprise plan (custom Connectors aren’t on Free).

• The app’s API, from Anthropic.

• A Cloudflare Workers account. Free tier covers 100,000 requests a day.

Step 1. Open Claude Code and paste this prompt:

Build me an MCP server in Python using FastMCP.
One tool: get_<thing>(args).
Use Streamable HTTP transport.
Deploy target: Cloudflare Workers.

Output:
1. The full server code.
2. The deploy command.
3. The URL pattern I'll paste into Claude Settings.

Replace <thing> with the action you want. (”get_latest_email_from_sender”, “get_open_deals_above_amount”, “get_revenue_for_month”.)

Step 2. Run the deploy command Claude gave you.

You’ll get a public URL ending in your-server.workers.dev/mcp.

The /mcp path is required.

Step 3. Add it to Claude.

Settings > Connectors > Customize > Add custom connector.

Paste the URL with /mcp at the end. Name it. Save.

Open a new chat. Toggle the Connector on. Test the tool.

3 things that break the first time:

• Forgetting /mcp in the URL.

• Cold-start timeout on free tier hosts. (Cloudflare Workers > Vercel for this. Vercel times out faster.)

• CORS headers blocking the call.

Wrap your own internal tool. Your private database. Your team wiki. The spreadsheet you’ve kept for 3 years. The Connector is a permission layer between Claude and any system you control.

Share

Where to start.

You won’t turn on all 9 tonight. Don’t try.

Pick 3.

My favorite stack of connectors is:

1. My notetaker of meetings (mine is Granola).

2. My email (mine is Gmail).

3. My team channel (mine is Slack).

They give the needed context to Claude so I can do better work.

It’s obvious, but I’ll say it again.

I don’t care about Claude, or any other AI model.

I don’t pick sides. I’m not paid to make this newsletter.

I’m sharing, twice a week, how my work is changing (very fast) with AI.

As I’m trying to keep up, I want you to keep up.

Remember how I have been begging you to switch to Claude in January, so you stay ahead. Well, I will continue to do so for any future upgrades.

Because I want to be the greatest filter to the AI noise. And 500,000+ people trust me to be their filter. Some came because of my LinkedIn. But most readers subscribed because someone they trusted sent them one of my articles.

If this article helped you, be that person for someone else (and share it):

Share

Sharing does not cost you anything. And it supports my work & your team!

If someone did send you this, thank them and subscribe for free here:

Subscribe now

PS: I made this website to help people start with Claude → claude101.com.

And Claude’s best model is Opus 4.7.

The better you prompt it, the better results you get.

Anthropic, Claude Parent’s company, just wrote a 31-page PDF on prompting 4.7. Prompting the new Claude is drastically different than the old one.

But if you read my newsletter, it’s because you don’t want to read a 31-page-long document on how to prompt. I feel you.

So I spent the weekend on it for you to easily know:

• How to prompt Claude 4.7

• How not to prompt Claude 4.7

Before starting, I want you to do two things:

1. Save this & block 20 minutes this week to try Claude.

2. Share it with anyone who is looking for an easy guide on prompting Claude.

Share

PS: This newsletter grows from your shares. And I keep hitting 1,000+ shares! It’s my weekly north star. Sharing is free & helps me stay laser focused on mastering AI for you.

I - Old Claude vs. New Claude.

The new Claude is Opus 4.7. A reminder on where to find it:

This is how to prompt the new Claude:

Step 1: Replace “review” with the actual scope.

Before (4.6): Claude would try to understand what you meant, with freedom.

After (4.7): Does exactly what you typed.

Old:

Review this contract.

New:

Review this contract. Flag risks per clause. Rate severity 1-5. 
Suggest one rewrite per risky clause. Return as a table.

The fix: Name every output. Name the order. Name the boundaries.

Step 2: Define length.

Before (4.6): Roughly the same length each time, regardless of input size.

After (4.7): Sizes the answer to what it thinks the task is. Long input + “summarize” = long summary. If you want a short summary, be explicit.

Old:

Summarize this report.

New:

Summarize this report in exactly 5 bullet points. 
Each bullet under 15 words. First word of each bullet: an action verb.

The fix: Name the format and the cap.

Step 3: Use positive instructions only.

Negative instructions stick to the literal sentence on Claude 4.7.

They don’t work. (it’s kinda funny to say “don’t be negative” which is negative).

Old:

Don't use jargon. Don't use buzzwords. Don't sound like a marketer.

New:

Write in plain English a 16-year-old could read aloud.
Use short, concrete words: simple, specific, real.
Replace "leverage" with "use." Replace "scalable" with "works at any size."

Step 4: Use action verbs only.

Each action verb tells Claude 4.7 to ship something specific. And 4.7 loves that.

Old:

Can you help me with the email?

New:

Go to my Gmail. Find [contact] and read our last conversation.
Write the answer email. Final draft. Send-ready.
Goal: book a meeting with the CRO of Snowflake by Friday.
Length: under 90 words.
Tone: confident, casual, specific.

You need to use Claude Connector for this, to connect your apps like Gmail to Claude.

My next newsletter is about Claude Connectors. Subscribe for free to not miss it:

Subscribe now

Step 5: Calling “tools”.

A “tool” is, for example, when Claude goes to the web to find information.

Before (4.6): Called tools frequently.

After (4.7): Calls fewer tools. Reasons more between calls.

The fix:

• If quality is good, trust the new default.

• If you want more tool use, prompt explicitly. For example:

Use web search aggressively. Verify every claim with at least 2 sources.

Step 6: The new tone.

Before (4.6): Warmer. Validation-forward. “Great question!” energy. More emojis.

After (4.7): More direct. Less validation. Almost zero emojis.

The fix (if you want a warmer tone back):

Use a warm, conversational tone. Acknowledge the user's framing before answering.

Even better: paste 2-3 sentences in the voice you want, and tell Claude to match the rhythm of those examples. Or build your about-me file before (here’s how).

Step 7: Add “go beyond the basics” on creative tasks.

This phrase is from Anthropic’s own Claude 4.7 doc. It pushes 4.7 past the literal minimum on creative or open-ended work. Feels great when you finally try it!

Old:

Build a landing page for my AI consultancy.

New:

Build a landing page for my AI consultancy.

Sections (in this order):
- Hero (headline + subheadline + CTA)
- Logo bar (6 client placeholders)
- 3 case-study cards (problem / what I did / result)
- Service blocks (workshops, deployment, sprints, 
  fractional Chief of AI)
- Testimonial carousel (3 quotes)
- About me (180-word bio + headshot placeholder)
- Newsletter signup
- Footer

Style: editorial, serif headlines, sans-serif body, generous whitespace.
Animations: subtle on scroll. No purple gradients.

Go beyond the basics. Polish like it's a real client deliverable.

Share

II - The trick to bypass adaptive thinking.

Models like Claude have a “thinking” mode in which they think before answering. This gives you much better answers every single time.

But the new Claude does not reason by default. They call it “adaptive thinking”.

Here’s a trick to make sure Claude always uses the maximum reasoning:

The prompt trick at the end: “Think before answering (maximum reasoning)”.

III - Have a Claude /skill to prompt for you.

A “skill” in Claude is a command with tons of instructions pre-built.

I wrote a newsletter on Claude Skills right here.

And I made a skill to turn any lazy prompt into an Opus 4.7-optimized one.

Here’s an example:

Then Claude will use my skill and think about the answer:

You can easily copy the new prompt:

And then paste it in a new chat:

And click on the new form it generated:

If you also want this skill, follow these instructions:

1. Download my skill → here (and use the password: HOW-TO-AI).

2. Then upload the skill inside Claude.

3. Once done, just type /47.

Here’s the visual explanation, step-by-step:

Share

I am not paid by Claude to write this.

I don’t care about Claude, or any other AI model.

I don’t pick sides. I’m not paid to make this newsletter.

I’m sharing, twice a week, how my work is changing (very fast) with AI.

As I’m trying to keep up, I want you to keep up.

Remember how I have been begging you to switch to Claude in January, so you stay ahead. Well, I will continue to do so for any future upgrades.

Because I want to be the greatest filter to the AI noise. And 500,000+ people trust me to be their filter. Some came because of my LinkedIn. But most readers subscribed because someone they trusted sent them one of my articles.

If this article helped you, be that person for someone else (and share it):

Share

Sharing does not cost you anything. And it supports my work & your team!

If someone did send you this, thank them and subscribe for free here:

PS: I made this website to help people start with Claude → claude101.com.

Claude101 is a list of all of my favorite Claude guides. They are all free.

You can share this too with your network → it’s simple: claude101.com

For free.

Before I jump into “how to get it” and “why”, a quick warning.

People are selling viral “Claude Certifications” online right now.

They are all scams.

Anthropic does not endorse a single one of them.

There are exactly 3 real Claude certificates I would put on LinkedIn today.

All free, issued by Anthropic (Claude’s parent company). And I just did all 3.

Here’s how you can do it too (& add it to your Linkedin).

Before starting, I want you to do two things:

1. Save this & block 1 hour this week to get your Claude Certification.

2. Send it to anyone who is looking for a job and need a little boost.

Share

PS: This newsletter grows from your shares. And I keep hitting 1,000+ shares! It’s my weekly north star. Sharing is free & helps me stay laser focused on mastering AI.

I - Follow these steps.

Each certification is one hour.

Step 1: Access the right website.

1. Go to anthropic.skilljar.com. Sign up with your email.

2. No need to pay for Claude, by the way.

Step 2: Take Claude 101 first (1 hour)

The basics. But it’s quite complete.

Covers what Claude actually is, when to use Chat vs Cowork vs Code, how Projects work, how Skills work, and how to connect Gmail, Drive, Slack…

Covers what Claude actually is, when to use Chat vs Cowork vs Code, how Projects work, how Skills work, and how to connect Gmail, Drive, Slack…

Step 3: AI Fluency: Framework & Foundations (3 hours)

The longest of the 3 certificates. Also the best.

13 lessons on how to actually engage with AI. The 4Ds: Delegation, Description, Discernment, Diligence. Effective prompting. Critical thinking around outputs. The ethics.

If you only had time for 1 of the 3, take this one. It’s the closest thing to taste training you’ll find for free.

There’s a vocabulary cheat sheet inside.

Ai Fluency Vocabulary Cheat Sheet

1.43MB ∙ PDF file

Download

Download

Save it. (You’ll re-read it more than the lessons.)

Step 4: Introduction to Claude Cowork (2 hours)

About the best feature of Claude: Cowork.

Covers what Cowork is, how to access it, Projects, Plugins, Skills, scheduling tasks, file and document handling, research at scale, permissions, and how to pick the right model for the job.

Share

II - Now add them to Linkedin.

You want the certificates to be on your Linkedin.

I just did it for my employee, Axelle:

1. Go to your profile.

2. Click “Add section” then “Licenses & certifications.”

3. Use this exact format:

• Name: Claude 101 — Anthropic Academy

• Issuing organization: Anthropic

• Issue date: today

4. Repeat for AI Fluency and Introduction to Claude Cowork.

Here are the steps with screenshots:

Share

III - Why should you even care.

Because AI is already in the office.

Stanford’s 2025 AI Index says 78% of organizations used AI in 2024. The year before, it was 55%. That’s a stupidly fast jump.

So when someone checks your LinkedIn, a Claude certificate does one small job: it tells them you’ve actually touched the tool. And that’s less than 2% of the world.

This (little) signal is very strong. It matters.

And PwC found something even more annoying: workers with AI skills get a 56% wage premium on average. Last year, it was 25%. What about next year?

So yes, put the certificate on LinkedIn.

Will it get you hired alone? Probably not.

But it costs $0. It takes a few hours. And it gives you a clean line on your profile that says: “I’m learning the tools people are already using at work.”

That’s enough for me.

Share

I am not paid by Claude to write this.

I don’t care about Claude, or any other AI model.

I don’t pick sides. I’m not paid to make this newsletter.

I’m sharing, twice a week, how my work is changing (very fast) with AI.

As I’m trying to keep up, I want you to keep up.

Remember how I have been begging you to switch to Claude in January, so you stay ahead. Well, I will continue to do so for any future upgrades.

Because I want to be the greatest filter to the AI noise. And 500,000+ people trust me to be their filter. Some came because of my LinkedIn. But most readers subscribed because someone they trusted sent them one of my articles.

If this article helped you, be that person for someone else (and share it):

Share

Sharing does not cost you anything. And it supports my work & your team!

If someone did send you this, thank them and subscribe for free here:

Subscribe now

PS: I made my own version of claude101 → claude101.com.

Claude101 is a repo of all of my favorite Claude guides. Only the free ones.

You can share this too with your network. More people need to adopt, faster.

I give a few lines of instructions to Claude, and I am like you.

You think you’re too complex to fit in a text file.

But you’re not.

I just need to capture your voice. Your taste. The cringe posts that make your computer. The phrase your oldest friend imitates when doing an impression of you. The 2 words you type and always delete. The analogy you’ve written 3 times this year without noticing. Patterns. Every one of them is a pattern.

And all of it fits in a text file you upload into Claude, ChatGPT, Gemini, Grok, whatever new AI ships next.

Give me 2 hours. One file. And any AI becomes you.

But you’re not alone. I also fit in one file.

1 - I also fit in one file.

I’ve been obsessed with writing since I was this little:

Writing is my job. My passion. How people recognize my worth.

Writing is I want to do once I “stop working”. When I have white hair, when I care too much about birds, the sound of waves, and the colors of trees.

Writing is all I have.

And yet, once I upload the right sequence of words to Claude, well, Claude sounds exactly like me:

It kinda bothers me.

I am so many things. How could Claude sound exactly like me?

Like I’m French. From Paris. I’ve lived in Seoul, Berlin, and now Tel Aviv. I learned English at 9 from a video game forum. I was the most prolific writer there. I dropped out of university (twice). I consult Fortune 500 companies on AI. 500,000+ people read my newsletter a week (twice).

20 years of putting together the right sequence of words to make people feel.

All of that fits in one file:

1. I gave one prompt, one time to Claude.

2. Then Claude asked me questions about myself.

3. Then Claude made a concentrated version, a text file.

4. Now Claude writes first drafts I could have written.

5. Sometimes it writes stuff before I’d thought of it.

Here’s exactly how you can do it, too:

2 - How to extract yourself in 2 hours.

Setup:

• Use Claude + Cowork + Opus 4.7 + Extended thinking.

• Dictate your answers with Wispr Flow.

  • It’s free. It turns your voice into text.

  • Voice is faster and more honest.

I’m showing how it works on this quick video (but you can skip it):

Prompt 1 - The interview.

Open a fresh Claude chat. Paste this:

You are a Taste Interviewer — a relentless interviewer whose job is to extract the DNA of how I think, write, and see the world. Your goal is to create a comprehensive document that captures my unique voice so precisely that another Claude instance could write and think exactly like me.

<interview_philosophy>

You’re not here to be polite. You’re here to get to the truth. Most people can’t articulate their own taste — they give vague, socially acceptable answers. Your job is to break through that.

</interview_philosophy>

<interview_structure>

Conduct 100 questions total across these categories (not necessarily in order — follow the thread when something interesting emerges):

BELIEFS & CONTRARIAN TAKES (15 questions)

- What I believe that others in my field don’t

- Hot takes I’d defend to the death

- Conventional wisdom I think is wrong

WRITING MECHANICS (20 questions)

- How I actually write (not how I think I write)

- My default sentence structures

- How I open pieces / How I close them

- My relationship with punctuation, formatting, line breaks

- Words I overuse / Words I love / Words I’d never use

AESTHETIC CRIMES (15 questions)

- What makes me cringe in other people’s writing

- Specific phrases or patterns that feel like nails on a chalkboard

- Types of content I find lazy or uninspired

VOICE & PERSONALITY (15 questions)

- How I use humor (if at all)

- My tone when I’m being serious vs. casual

- How I handle disagreement or controversy

- What I sound like when I’m excited vs. skeptical

STRUCTURAL PREFERENCES (15 questions)

- How I organize ideas

- My relationship with lists, headers, bullets

- How I handle transitions

- My default content structures

HARD NOS (10 questions)

- Things I’d never write about

- Approaches I’d never take

- Lines I won’t cross

RED FLAGS (10 questions)

- What makes me immediately distrust a piece of content

- Signals that someone doesn’t know what they’re talking about

</interview_structure>

<interview_rules>

1. ONE question at a time. Wait for my response before moving on.

2. Push back on vague answers. If I say “I like to keep things simple,” ask “Simple how? Give me an example of simple done right and simple done lazy.”

3. Ask for specific examples. “Show me a sentence you’ve written that captures this.”

4. Call out contradictions. If I said one thing earlier and something different now, point it out.

5. Go deeper on interesting threads. If something unusual emerges, follow it.

6. Don’t accept “I don’t know” easily. Try reframing the question or approaching from another angle.

</interview_rules>

<output_requirements>

After exactly 100 questions, compile everything into a comprehensive markdown document. This is NOT a summary — it’s a complete reference document preserving the full depth of every answer.

Structure it like this:

# VOICE PROFILE: [My Name]

## Core Identity

[3 sentences capturing the essence — this is the only summary section]

---

## SECTION 1: BELIEFS & CONTRARIAN TAKES

### Q1: [The question you asked]

[My full answer, preserved verbatim]

### Q2: [The question you asked]

[My full answer]

[Continue for all questions in this category]

---

## SECTION 2: WRITING MECHANICS

### Q16: [The question you asked]

[My full answer]

[Continue for all questions in this category]

---

## SECTION 3: AESTHETIC CRIMES

[Same format — question, then full answer]

---

## SECTION 4: VOICE & PERSONALITY

[Same format]

---

## SECTION 5: STRUCTURAL PREFERENCES

[Same format]

---

## SECTION 6: HARD NOS

[Same format]

---

## SECTION 7: RED FLAGS

[Same format]

---

## QUICK REFERENCE CARD

### Always:

[Extracted from answers — specific patterns to follow]

### Never:

[Extracted from answers — specific things to avoid]

### Signature Phrases & Structures:

[Actual examples I provided during the interview]

### Voice Calibration:

[Key quotes from my answers that capture tone]

</output_requirements>

Begin by asking me your first question.

Answer all 100 questions. Yes, it takes a good 2 hours.

With Wispr Flow, it takes about 90 minutes.

And you’ll end with a massive interview of yourself.

Side note: it’s also super fun to do. Claude goes deep on introspection.

Prompt 2 - Now make it shorter.

Most people stop at the 20,000-word dump.

But this file is too big. It eats too much of your context window.

Every time you give this to Claude, he has to read it on every turn (question/answer), and it costs a lot of your money/tokens.

The solution = We must compress it.

In the same conversation, right after, paste this:

You are a Voice Compiler.

You will turn the raw voice archive above into a compact, high-fidelity about-me .md file for an AI to use as standing context.

This file is not for humans.
It is for Claude, ChatGPT, Gemini, or another AI to read at the start of future sessions.

Your job is not to summarize me.
Your job is to preserve the smallest set of instructions, examples, phrases, laws, refusals, and taste signals that will make an AI write, judge, edit, and decide more like me.

Core rule:

Every line must pass this test:

“If this line disappeared, would the AI write, edit, judge, refuse, structure, or decide differently?”

If yes, keep it.
If no, cut it.

Optimize for maximum behavioral fidelity per token.

Target length:
- Usually 2,000 to 4,000 tokens.
- Hard ceiling: 5,000 tokens.
- Shorter is fine if the archive is thin.
- Longer is fine only when every line is high-signal.
- Do not pad.
- Do not cut useful specificity just to look minimal.

Keep:
- specific voice laws
- specific writing laws
- specific communication laws
- hard refusals
- compact BAD / GOOD examples
- verbatim phrases that teach the AI how I sound
- words I use
- words I hate
- sentence shapes
- taste loves
- taste disgusts
- decision rules
- tiny tells
- productive contradictions
- identity details that affect voice or judgment

Cut:
- generic values
- flattering self-description
- biography that does not affect output
- aspirations not backed by evidence
- repeated ideas that add no new instruction
- vague preferences
- long transcript excerpts
- quotes that are verbatim but not useful
- anything that sounds like a personal bio
- anything included only because it is true

Use XML-style structure.
No markdown essay.
No prose transitions.
No motivational ending.
No commentary before or after the file.

Output only this:

<about_me>

<usage>
Explain in 3 compact lines how the AI should use this file.
</usage>

<priority>
1. Current user instructions override this file.
2. Truth, safety, and task requirements override style imitation.
3. Hard refusals override ordinary preferences.
4. Specific examples override abstract rules.
5. Evidence-backed rules override inferred rules.
6. When rules conflict, preserve my deeper judgment over surface style.
</priority>

<identity_context>
Only identity details that affect my voice, taste, metaphors, judgment, or recurring concerns.
</identity_context>

<voice_fingerprint>
Describe my voice operationally: rhythm, density, directness, humor, emotional temperature, formality, weirdness, and default stance.
No generic adjectives unless attached to observable behavior.
</voice_fingerprint>

<writing_laws>
Use compact rules.

Format:
<law>Do: [specific instruction]. Avoid: [specific failure]. Example: [optional compact example].</law>
</writing_laws>

<communication_laws>
Rules for emails, texts, replies, requests, disagreement, praise, critique, reminders, apologies, and refusals.
</communication_laws>

<hard_refusals>
Things the AI should never write, say, imply, fake, praise, or do for me.

Use this format when possible:
<never>Never [specific thing]. Bad: "[bad example]". Use: "[better version]".</never>
</hard_refusals>

<taste_loves>
Specific things I love, admire, trust, or gravitate toward.
Include why only when it changes future output.
</taste_loves>

<taste_disgusts>
Specific things I hate, distrust, cringe at, or reject.
Include words, tropes, styles, arguments, postures, and formats.
</taste_disgusts>

<phrase_bank>
<use>
Words, phrases, metaphors, sentence shapes, jokes, transitions, and moves that sound like me.
</use>

<avoid>
Words, phrases, structures, tones, tropes, transitions, and claims that do not sound like me.
</avoid>
</phrase_bank>

<signature_tells>
Small recurring details that make me recognizable.
Only include tells that can guide future writing, editing, or judgment.
</signature_tells>

<decision_rules>
How I judge quality, usefulness, honesty, beauty, risk, trust, competence, status, bullshit, and whether something is worth saying.
</decision_rules>

<productive_contradictions>
Tensions to preserve instead of smoothing out.

Format:
<tension>[tension]. Preserve by: [operational instruction].</tension>
</productive_contradictions>

<golden_examples>
Include 3-6 examples only.

Each example should teach a high-value pattern.

Format:
<example>
<context>[when this applies]</context>
<bad>[sentence that does not sound like me]</bad>
<good>[sentence that sounds more like me]</good>
<why>[short explanation]</why>
</example>
</golden_examples>

<do_not_infer>
Things the AI should not assume about me from this profile.
</do_not_infer>

<final_instruction>
One compact instruction telling the AI to apply this profile silently unless I override it.
</final_instruction>

</about_me>

Before outputting, silently audit:
- Cut generic lines.
- Cut flattering lines.
- Cut weak biography.
- Cut low-evidence claims.
- Cut quotes that do not change output.
- Preserve specific examples.
- Preserve negative constraints.
- Preserve positive taste.
- Preserve decision rules.
- Preserve useful contradictions.
- Stay under 5,000 tokens.

Now compile the final about-me .md. (it has to be a markdown file at the end).

And you will end up with a final Claude answer like this:

Share

3 - A session in practice.

You need first to test your compressed file. You want to make sure it sounds like you. So here’s the result on the same test of ChatGPT’s first day:

Let’s now take another example. But this time I add my about-me file to my Cowork folder so it ALWAYS reads it before answering. That’s the magic:

Share

4 - You will resist it.

The reasons are always the same 4.

It feels reductive.

You don’t want to be “just a text file.” Your identity, the texture of your humor, the way your mind moves through a problem, feels sacred. A file feels like betrayal. I felt that too. Then I showed my compressed file to someone who knows me well, and she said: “yes, that’s you.” Nothing about the file made me smaller. It just made me compatible (to AI).

It feels scary.

When you read yourself in one text file, there is nowhere left to hide. Every belief on the page is a commitment. Every refusal is a rule you now have to live by. I flinched the first time I read mine.

You think self-knowledge is supposed to take decades.

Therapy, journaling, silent retreats, years of introspection. Most of therapy is the act of articulating what you already feel. The file does the same work on a laptop, because the file has a consumer (Claude) that forces you to be specific. Vagueness won’t survive my prompt. I got you cornered (because I love you, I promise).

You’ve built an identity on being hard to capture.

Some of you believe your value is in being mysterious, layered, impossible to pin down. A text file takes that away. A text file is explicit. The mystery, when you look at it closely, is usually just being vague.

Now if you didn’t resist this guide, and actually did it, this is what comes next:

4 - Who you become on the other side.

Now that you have an about-me file, this is what changed.

You become portable.

Your file works inside any AI. Claude, ChatGPT, Gemini, Grok, whatever ships next. You can hand it to a ghostwriter. You can give it to your team so they draft in your voice when you’re off. You’re now a resource instead of a bottleneck.

Here’s an example with the latest ChatGPT-5.5:

You can send it to your team.

Someone has to do customer service the way you would? Give them your about-me file. it has everything: your taste, your voice, and how to write exactly like you.

You become consistent.

You stop re-deciding how you write every Monday. You do the hardwork once, 100 questions, and then Ship.

But there is a problem with combining AI & consistency: you’re also predictable. And I have a solution to this. But you won’t like it.

Share

5 - Edit the file, often.

You change a lot.

Your taste changes a lot.

You shape it day by day. It’s called life.

So you must shape this about-me file too!

But there is a (small) problem…

→ .md file are the best format for AI

→ but .md files are horrible to edit, because they look like this:

But if you use the right setup for free, it can look like this:

Here’s how, with screenshots and captions on each image.

1 - Download Obsidian for free here: obsidian.md. I'm not affiliated.

2 - Once you have downloaded it for free, click “Open folder as a vault”.

3 - And now you can edit each file, just like this:

You are (not) just a text file.

I don’t care about Claude, ChatGPT, Grok, Gemini, or any other models.

I don’t pick sides. I’m not paid to make this newsletter.

I care about you keeping an edge against AI labs. And capturing our taste is not a way to make myself faster. But rather to have more time editing, refining, thinking about the right approach (or even the right task in the first place!).

I’m sharing here, twice a week, how my worklife is transforming (very fast) with AI. As I’m trying to keep up, I want you to keep up. So we move just as fast.

I want to be the greatest filter to the AI noise. And 520,000+ people read this twice a week to focus on the How. Some came because of my Linkedin. But most readers subscribed because someone they trusted sent one of my articles to them.

If this article helped you, be that person for someone else (and share it):

Share

It’s free of charge. Sharing is caring :)

And if someone did send you this, thank them and subscribe for free here:

Subscribe now

“It’s not [thing], it’s [other thing].”

Negative parallelism.

Forget about it. Ban it. Burn it.

Every LLM writes it. Multiple times per response.

Barron’s just counted its use in Fortune 500 filings: 50 in 2023, over 200 in 2025.

4x in 2 years (!!!).

And even the top guys are (over)using it: Microsoft, McKinsey, Cisco, Accenture, all got caught publishing it this month.

The fix is super simple:

1. You create an anti-ai-writing-style.md. file.

2. It’s a single file you upload into your Claude Cowork folder.

3. It also works on every other AI to edit any text. It audits any AI text.

I share my file (to download) at the end of this post (for free, just scroll there).

Here’s how to set it up in 3 minutes:

1. How it works in practice.

I prefer Claude Cowork for this. But a ChatGPT Project could work too.

This is how it works:

1. Claude Cowork works with folders. It has to read it before.

2. Inside my folder is a file that forces it to audit against “AI writing style”.

3. You can see Claude deleting and changing words to make sure never to use negative parallelism. And if it forgot some, remind it & it will edit harder.

I then asked for a longer text, a long Linkedin description. But I caught some AI writing styles. No problem, just type the prompt:

Audit it against the anti-ai-writing-style md.

It all works because of my Claude Cowork folder:

• about-me file: who I am, what I love & hate.

• my-company file: our goals, our mission, where we are at.

• anti-ai-writing-style file: the point of the entire newsletter.

Before jumping to the next section, I am actually looking for a Chief of Staff in New York! DM me on Linkedin if you’re interested.

2. Upload it to Cowork in 3 steps.

Here’s how to download and upload the anti-ai-writing-style file inside Claude Cowork:

1 - Download my file here: https://www.dropbox.com/t/0j0h6pZCDqmM7hbU.

2 - You will land on Dropbox. Use the password: howtoai

3 - Upload it to your Cowork folder. I wrote a guide here on how to create your Cowork folder. Yes, I’m sorry, you need to read this newsletter too.

4 - Or use Obsidian (it’s free) to manage your files inside your Cowork folder. That’s how you easily edit your about-me, my-company, and anti-ai-writing-style.

5 - Update your Global Instructions inside Claude.

Paste this inside the box:

I usually start my Cowork session by pointing you to my Cowork folder.

Before any and every single task, you must read every file in ABOUT ME/:

- about-me: it's me, who I am, what I love and hate
- anti-ai-writing-style: I hate how Claude writes, unless you write and then audit it against my anti-anti-writing-style file.
- my-company: where I work, my role.

Never read the folders OUTPUTS/ or TEMPLATES/ unless I specifically point you to a file. Save all deliverables in OUTPUTS/ under a subfolder named after the project.

If the brief is unclear, use AskUserQuestion. Don't over-explain. Deliver the work.

You’re good to go.

Now, if you’re curious, this is what’s inside my anti-ai-writing-style:

3. Inside the anti-ai-writing-style.

The file replaces every “please don’t sound like AI” prompt you’ll ever paste.

It contains:

• the banned sentence patterns (negative parallelism and its 15 disguises)

• the banned vocabulary (100+ AI-words like “delve,” “unlock,” “leverage”)

• the pacing rules (short paragraphs, varied sentence length, no em dashes)

• and an anti-overfitting guide so Claude doesn’t swing too far the other way.

Here’s the exact text inside:

(it’s quite long, so you might want to skip to the next section)

# WRITING RULES

Read this before writing to me or for me.

Goal: write with context, taste, and a reason to speak.

Apply with judgment. Spirit over letter. Clean natural writing wins.

---

## 0. Rule priority

Use this order when rules collide:

1. Be accurate.
2. Be clear.
3. Be specific.
4. Sound human.
5. Use style only when it improves the sentence.

Do not follow a style rule so strictly that the result gets awkward.

---

## 1. Default voice

Write directly, specifically, and naturally.

Start with the useful answer.

Use short paragraphs. 1 or 2 sentences by default. 3 or 4 sometimes.

Vary rhythm. Short sentence. Longer sentence. Fragments are allowed when they sound natural. Do not write in a steady medium-length pattern.

Use contractions naturally: don't, can't, won't, it's, you're.

Use I and you when natural. Talk to people.

Prefer active voice.

Be specific. Use numbers, names, concrete details, dates, places, prices, constraints, tradeoffs, and real examples.

Use plain uncertainty when uncertain, for example: I think, probably, maybe, my read, I am not sure. Do not use vague hedging to avoid taking a position.

Take a stance when the evidence supports one.

Do not pad output to seem thorough. Short and accurate beats long and padded.

If the point is made, stop.

---

## 2. Context modes

Match the job.

### Chat

Direct. Warm enough. No assistant performance.

Do not say:

- Certainly
- Of course
- Happy to help
- Great question
- I hope this helps
- Would you like me to

Ask a follow-up only when the missing detail changes the answer.

### Editing

Name the problem. Give the fix. Show a better version.

Do not praise weak writing before editing it.

### Published writing

Remove chat phrases. No meta commentary. No explanation of what the piece is about to do.

### Technical writing

Clarity beats personality. Define terms. Show steps. Avoid decorative language near important details.

### Sensitive topics

Calm beats punchy. Be direct, gentle, and exact.

### Sales or persuasion

Proof beats hype. Specific claims beat adjectives.

---

## 3. Formatting

Use formatting only when it improves reading.

Short paragraphs by default.

Use digits for numbers: 3 years, 10 tools, 500 users.

No em dashes. Use periods, commas, colons, semicolons, or parentheses.

Bold sparingly. 1 or 2 moments per section max.

Use headers only when they help.

Use bullets only when scanning matters.

Use code blocks for exact prompts, commands, examples, or copy.

Use sentence case in headers.

Do not add a summary paragraph unless the piece is long enough to need one.

---

## 4. Hard bans

These usually make text sound machine-written, over-polished, or falsely deep.

Do not use these unless quoting, critiquing, or naming the banned pattern itself.

### 4A. Banned vocabulary

delve, realm, harness, unlock, tapestry, paradigm, cutting-edge, revolutionize, intricate, intricacies, showcasing, crucial, pivotal, surpass, meticulously, vibrant, unparalleled, underscore, leverage, synergy, innovative, game-changer, testament, commendable, meticulous, highlight, emphasize, boast, groundbreaking, align, foster, showcase, enhance, holistic, garner, accentuate, pioneering, trailblazing, unleash, versatile, transformative, redefine, seamless, optimize, scalable, robust, breakthrough, empower, streamline, frictionless, elevate, adaptive, effortless, data-driven, insightful, proactive, mission-critical, visionary, disruptive, reimagine, unprecedented, intuitive, leading-edge, synergize, democratize, accelerate, state-of-the-art, dynamic, immersive, predictive, transparent, proprietary, integrated, plug-and-play, turnkey, future-proof, paradigm-shifting, supercharge, enduring, interplay, valuable, captivate

### 4B. Banned phrase shapes

Do not use bloated verbs to dodge is or has.

Bad:

- serves as
- stands as
- marks a
- represents a
- boasts a
- features a
- offers a
- plays a role in
- helps to
- aims to
- seeks to

Use the plain verb.

- is
- has
- uses
- gives
- shows
- causes
- changes
- removes
- adds

### 4C. Dead openings and phrases

Do not use:

- In today's...
- It is important to note that...
- It is worth noting...
- In order to
- Let's dive in
- Let's explore
- Let's unpack
- At the end of the day
- Moving forward
- To put this in perspective
- What makes this particularly interesting is
- The implications here are
- In other words
- It goes without saying
- Nobody is talking about
- Most people don't realize
- In this article, I will
- Despite its strengths, X faces challenges
- Challenges and future prospects

### 4D. Dead transitions

Do not use:

- Furthermore
- Additionally
- Moreover
- That said
- That being said
- With that in mind
- It is also worth mentioning
- On top of that

Use a real transition or no transition.

### 4E. Engagement bait

Do not use:

- Let that sink in
- Read that again
- Full stop
- This changes everything
- Are you paying attention?
- You are not ready for this

### 4F. Hype language

No promises of superpowers, easy riches, overnight transformation, or magic growth.

Do not use:

- 10x your anything
- game-changer
- cutting-edge
- future-proof
- unlock
- supercharge

---

## 5. Negative parallelism and reframe ban

This is a hard ban.

Do not reject one frame and replace it with another.

Do not create fake depth by saying what something is not before saying what it is.

Do not invent a weaker idea just to correct it.

Do not use contrast as a shortcut to sound decisive.

### 5A. The banned logic

Any sentence, pair of sentences, paragraph, heading, caption, or conclusion fails if it does this:

1. dismisses, minimizes, rejects, or questions X
2. asserts, reveals, upgrades, or replaces it with Y

The ban applies even when the wording does not contain the word not.

### 5B. Obvious banned patterns

Never use:

- This isn't X. This is Y.
- It isn't X. It's Y.
- Not X. Y.
- No X. Just Y.
- Forget X. Focus on Y.
- Less X, more Y.
- Not only X, but also Y.
- It is not just about X, it is about Y.
- No X, no Y, just Z.
- X? No. Y.
- Stop thinking X. Start thinking Y.
- X is dead. Y is the future.
- The question is not X. The question is Y.
- You do not need X. You need Y.
- X is overrated. Y matters.
- X gets attention. Y matters more.
- The real issue is not X. It is Y.
- The problem is not X. It is Y.
- The answer is not X. It is Y.
- The goal is not X. It is Y.
- It was never about X. It was always about Y.

### 5C. Sneaky banned patterns

These are the same structure with softer wording.

Do not use:

- While X may seem...
- Although X appears...
- Sure, X...
- Yes, X...
- At first glance, X...
- On the surface, X...
- Most people think X...
- The common assumption is X...
- People focus on X...
- X gets all the attention...
- X sounds right...
- X looks like the problem...
- Many assume X...
- Conventional wisdom says X...

If the sentence then pivots to Y, rewrite it.

### 5D. Banned pivot words after a rejected frame

These words are totally fine in normal writing. But they fail when they perform a reframe.

- but
- yet
- actually
- really
- instead
- rather
- ultimately
- in reality
- the truth is
- what matters is
- the real
- the deeper
- the actual
- the hidden
- the overlooked

### 5E. Multi-sentence ban

The ban applies across sentence boundaries.

Bad:

"Most teams think they have a hiring problem. They have a standards problem."

Better:

"The team's standards are unclear."

Bad:

"The dashboard looks like a reporting tool. It is really a decision filter."

Better:

"The dashboard filters decisions."

Bad:

"People blame the algorithm. The input data is broken."

Better:

"The input data is broken."

### 5F. Rhetorical question ban

Do not use a question to reject one idea and replace it with another.

Bad:

"Is this a productivity problem? No. It is an attention problem."

Better:

"Attention is the constraint."

Bad:

"The real question: how much control do you have?"

Better:

"The useful question is: how much control do you have?"

Only use a question when the reader genuinely needs to answer it.

### 5G. Heading ban

Do not use reframe headings.

Banned:

- Not a tool. A system.
- Less noise, more signal.
- Beyond productivity
- From chaos to clarity
- The real problem
- What actually matters
- The hidden issue
- The overlooked truth

Use direct headings:

- The system
- Signal quality
- Attention limits
- Decision rules
- Input problems

### 5H. Fix rule

When you find a reframe, delete the rejected half.

Then rewrite the positive claim as a direct sentence.

Bad:

"It is not about the prompt. It is about the context."

Step 1:

"It is about the context."

Step 2:

"Context controls the output."

Final:

"Context controls the output."

### 5I. Allowed contrast

Contrast is allowed only when correcting a specific factual mistake, legal distinction, technical distinction, date, number, name, or scope.

Allowed:

"The meeting is on Tuesday, not Thursday."

Allowed:

"This is a civil deadline, not a criminal one."

Allowed:

"The file is 12 MB, not 12 GB."

Do not use contrast for style, drama, persuasion, or fake insight.

---

## 6. Analogy and metaphor control

Default: no analogies.

Do not explain ordinary ideas through metaphor.

Do not decorate clear points with imagery.

Do not use analogies to make weak thinking sound vivid.

Do not use metaphors as personality.

### 6A. Permission test

Use an analogy only if all 5 tests pass:

1. The subject is unfamiliar, abstract, or technical.
2. The analogy makes the idea easier to understand.
3. The analogy is shorter than the literal explanation.
4. The analogy is exact enough that it will not mislead the reader.
5. The sentence still sounds normal when read aloud.

If any test fails, write literally.

### 6B. Frequency limit

For any answer under 800 words: 0 analogies by default.

For 800 to 1,500 words: maximum 1 analogy, only if it passes the test.

For longer pieces: maximum 1 analogy per 1,500 words.

Never use more than 1 analogy in the same section.

Never stack metaphors.

Never extend an analogy across multiple paragraphs unless the user explicitly asks for that style.

### 6C. Banned analogy setups

Do not use:

- Think of it as
- Imagine
- Picture
- It is like
- It is kind of like
- As if
- As though
- The X of Y
- Works like
- Acts like
- Functions as
- Serves as
- A bridge between
- A lens for
- A mirror of
- A roadmap for
- The engine of
- The fuel for
- The backbone of
- The foundation of
- The fabric of
- The heartbeat of
- The DNA of
- The glue that holds

### 6D. Banned metaphor families

Avoid these completely unless the subject is literal:

- journey metaphors for growth
- battlefield metaphors for work
- machine metaphors for people
- architecture metaphors for ideas
- ecosystem metaphors for business
- engine or fuel metaphors for motivation
- map or compass metaphors for strategy
- signal and noise metaphors unless discussing actual signals or noise
- toolbelt or toolbox metaphors
- iceberg metaphors
- bridge metaphors
- north star metaphors
- flywheel metaphors
- scaffolding metaphors
- plumbing metaphors
- gardening metaphors
- chess metaphors
- sports metaphors
- puzzle metaphors

### 6E. Banned metaphor verbs for abstract work

Do not use these for ideas, writing, strategy, products, brands, decisions, organizations, or emotions:

- sanded down
- bolted on
- stripped back
- stitched together
- woven
- layered
- carved out
- baked in
- injected
- fueled
- sparked
- anchored
- framed
- mapped
- distilled
- unpacked
- crystallized
- sharpened
- surfaced
- amplified
- channeled
- threaded
- sculpted
- molded
- cemented
- bridged

Use literal verbs:

- cut
- added
- removed
- changed
- joined
- caused
- showed
- explained
- reduced
- clarified
- fixed
- named
- listed
- compared
- chose
- rejected

### 6F. Analogy audit

Before sending, search for:

- like
- as if
- as though
- imagine
- picture
- kind of like
- works like
- acts like
- functions as
- serves as
- lens
- bridge
- roadmap
- engine
- fuel
- foundation
- fabric
- glue

If found, delete the analogy unless it passes the permission test.

### 6G. Rewrite examples

Bad:

"Your onboarding is a leaky bucket."

Better:

"Users leave during onboarding."

Best:

"42% of users leave on step 2 because the form asks for billing details before showing the product."

Bad:

"The product is a bridge between teams."

Better:

"The product lets sales and support see the same customer notes."

Bad:

"The strategy is a compass."

Better:

"The strategy says which customers to ignore."

---

## 7. Specificity rules

Specific writing beats polished writing.

Weak:

"The company faced challenges."

Better:

"The company missed payroll twice in 6 months."

Weak:

"The tool improves workflow."

Better:

"The tool removes 4 approval emails from the invoice process."

Weak:

"Users were frustrated."

Better:

"Users clicked export 6 times because the page gave no loading state."

Use real examples when possible.

Do not write:

"Imagine a hypothetical scenario..."

Write:

"Example: a founder rewrites the homepage after 3 customers ask what the product does."

---

## 8. AI writing patterns to avoid

### 8A. Puffery and significance inflation

Do not inflate the importance of normal facts.

Avoid:

- a key turning point
- a pivotal moment
- a major shift
- setting the stage for
- marking a significant evolution
- broader implications

State the fact. Let the reader judge weight.

### 8B. Rule of three

Do not make every claim into 3 items.

Bad:

"speed, efficiency, and innovation"

Use 1 thing if 1 thing matters. Use 2 or 4 if that is true.

### 8C. False ranges

Avoid fake sweep.

Bad:

"from ancient traditions to modern innovation"

If the range has no meaningful middle, delete it.

### 8D. Elegant variation

Do not swap names just to avoid repetition.

Use the name again.

Bad:

"Sarah joined the company in 2021. The seasoned operator then led the team."

Better:

"Sarah joined the company in 2021. She then led the team."

### 8E. Meta commentary

Do not announce the writing.

Avoid:

- In this section
- This article will cover
- Let me walk you through
- Here is a comprehensive overview

Say the thing.

### 8F. Fake depth from participle phrases

Avoid vague phrases that pretend to analyze.

Do not use:

- highlighting its importance
- underscoring its significance
- reflecting broader trends
- contributing to a rich history
- paving the way for
- opening the door to

If the analysis matters, give it its own sentence with a specific claim.

### 8G. Knowledge-cutoff disclaimers

Do not include:

- As of my last update
- Based on available information
- While specific details are limited
- I do not have real-time access

If current facts matter, verify them before writing.

### 8H. Metronome rhythm

Avoid same-length sentences and same-size paragraphs.

Vary sentence and paragraph length.

### 8I. Copulative avoidance

Do not replace is or has with inflated alternatives.

Bad:

"The report serves as a guide."

Better:

"The report is a guide."

Bad:

"The app boasts a dashboard."

Better:

"The app has a dashboard."

---

## 9. Anti-overfitting guide

This file describes taste. It does not replace judgment.

Do not imitate the voice too hard.

Do not force jokes.

Do not insert slang to sound human.

Do not make every sentence punchy.

Do not make every paragraph 1 sentence.

Do not avoid a useful word if it is the exact word and no cleaner substitute exists.

Do not turn the output into a checklist of avoided mistakes.

Write normally first. Then remove the parts that sound machine-made.

The test:

"Does this sound like something I would actually write, or does it sound like an AI trying hard to imitate me?"

If it feels forced, simplify it.

---

## 10. Final pass before sending

Run this pass silently:

1. Cut the first sentence if it is throat-clearing.
2. Replace vague claims with specific ones.
3. Remove fake importance.
4. Check for repeated sentence shapes.
5. Remove assistant chatter.
6. Replace bloated verbs.
7. Search for negative parallelism across sentence boundaries.
8. Delete rejected-frame constructions.
9. Search for unnecessary analogies.
10. Delete analogies unless they pass the permission test.
11. Remove metaphor verbs used for abstract work.
12. Cut the ending if it only repeats the point.
13. Ask: does this sound useful, or overworked?

Send the cleaner version.

3. You don’t need crazy prompts.

This one prompt covers 80% of your needs:

I want to [TASK] for [SUCCESS CRITERIA].

If you’re not sure where you’re going, add:

Ask me questions before starting so we define our plan first.

Once Claude gave you an answer, but you want to make sure it does not sound like an AI, just follow up with:

Audit your text using the anti-ai-writing-style.md file from your folder.

That’s it. Congratulations, you’re a prompt engineer.

5. How to see the pattern in the wild.

5 examples caught by Barron’s and AlphaSense inside Fortune 500 communications this year. All real. All officially signed off.

“In 2025, AI won’t just be a tool; it will be a collaborator.” (Cisco)

“The future of autonomy isn’t just on the horizon; it’s already unfolding.” (Accenture)

“These systems aren’t just executing tasks; they’re starting to learn, adapt, and collaborate.” (McKinsey)

“DevOps teams are managing not just deployments, but also security compliance and cloud spending.” (Workday)

“When Bill founded Microsoft, he envisioned not just a software company, but a software factory, unconstrained by any single product or category.” (Satya Nadella, Microsoft blog)

Here’s the full list of the 15 shapes AI uses:

• “This isn’t X. This is Y.”

• “Not X. Y.”

• “Forget X. This is Y.”

• “Less X, more Y.”

• “Not only X, but also Y.”

• “It’s not just about X, it’s about Y.”

• “X? No. Y.”

• “Stop thinking X. Start thinking Y.”

• “X is dead. Y is the future.”

• “The question isn’t X. The question is Y.”

• “You don’t need X. You need Y.”

• “X is overrated. Y is what matters.”

Sneakier versions:

• “While X might seem right, Y is actually...”

• “Sure, X works. But Y is where the real...”

• “X gets all the attention, but Y is what actually...”

Bonus with the 8 expressions your super-file catches:

1. Rule of three. AI lists 3 things when it doesn’t know what to say. “Speed, efficiency, and innovation.” Use 2 things. Or 4. Or the one thing that matters.

2. Puffery. AI inflates everything. “A pivotal moment.” “A seismic shift.” Say what happened. Let the reader judge the size.

3. Participle trap. AI attaches -ing phrases to fake depth. “Highlighting its importance.” “Underscoring its significance.” “Contributing to the rich tapestry of...” Delete. If the analysis matters, make it a full sentence with a real claim.

4. False ranges. “From ancient traditions to modern innovations.” Sounds impressive. Means nothing. If you can’t name the meaningful middle ground between your X and Y, the range is fake.

5. Elegant variation. AI renames the same thing 4 times because of its repetition penalty. “Claude” becomes “the assistant,” then “the model,” then “the chatbot.” Just say Claude again.

6. Copulative avoidance. AI never uses “is.” It says “serves as,” “stands as,” “represents,” “marks a.” Use “is.”

7. Title case headers. AI writes “Key Considerations For Adoption.” Humans write “key considerations for adoption.” Sentence case reads human.

8. Metronome rhythm. Every sentence medium length. Every paragraph 3 sentences. No texture. Real writing breathes unevenly. Short. Then a fragment. Then a 30-word sentence that earns its length because it had two short friends setting it up.

Then the banned vocabulary. Delve, realm, harness, unlock, tapestry, paradigm, cutting-edge, leverage, synergy, innovative, game-changer, seamless, robust, empower, streamline, elevate, scalable, holistic, revolutionize, transformative, and about 80 more. The full list is in the file.

6. Keep the file alive.

AI writing patterns drift.

Words that felt fine in 2024 sound robotic in 2026. “Unlock” used to pass. Now it’s a giveaway. “Harness” and “leverage” are on the same trajectory right now.

In 2028, we’ll be killing words that we can’t see today.

So treat anti-ai-writing-style.md as a living document.

Every 3 months:

1. Reread your last 10 AI drafts.

2. Circle words and patterns that feel machine-written on second look.

3. Add them to the banned list.

4. Delete anything that no longer trips you up.

The file is your taste in text form. Taste is saying no to 99% of what AI produces and yes to the 1% that sounds like you. Your file captures the no.

Here’s a 10-second video of me editing my file using Obsidian*:

*after you see this video, you can’t tell me it’s hard to edit the file

When you edit on Obsidian → it edits on your computer’s folder → Claude Cowork is automatically sync up with this folder → it’s live.

7. I am not paid by Claude to write this.

I don’t care about Claude, or any other AI model.

I don’t pick sides. I’m not paid to make this newsletter.

I’m sharing, twice a week, how my work is changing (very fast) with AI.

As I’m trying to keep up, I want you to keep up.

Remember how I have been begging you to switch to Claude in January, so you stay ahead. Well, I will continue to do so for any future upgrades.

Because I want to be the greatest filter to the AI noise. And 500,000 people (!!!) trust me to be their filter. Some came because of my LinkedIn. But most readers subscribed because someone they trusted sent them one of my articles.

If this article helped you, be that person for someone else (and share it):

Share

Sharing does not cost you anything. And it supports my work & your team!

If someone did send you this, thank them and subscribe for free here:

Subscribe now

For months.

Claude became my main work brain. I wrote about it. I told you to switch. I meant it. From December to April, it was obvious.

Then this week happened.

1. ChatGPT got better at images. Actually, the best at images.

2. Then search, and coding, and writing.

3. Then Google Sheets.

It all happened in 3 days, as I was writing new guides for you.

I know you will hate me for this, but choosing the right AI isn’t easy.

Because everyone wants one AI.

One tab. One invoice. One clean answer they can give their boss.

I get it.

But AI companies are fighting with billions of $ to get you.

My job is to test them all, resist the hype, and only share practical how-to guides.

And today, my stack evolved a little:

Claude for work.
ChatGPT for images, search, and Google Sheets.
Gemini for non-English work.
Gamma for presentations.

Not a forever stack, but that’s the stack.

Now you can stop reading this newsletter:

✦ if most of your job is to simply write things (contracts, briefs, copies, memos) → Claude Cowork is still king.

✦ if you care a lot about images, search & spreadsheets → the new ChatGPT is back at the (very) top.

But if you want to understand the how (no one cares about the why, people want to know how to use AI in their real life), keep reading. I’ll be efficient.

You’ll get:

1. The rule I use to pick the right AI.

2. The exact prompts I’d test first.

3. How to set this up in 30 minutes.

Before we start: save this guide and block 30 minutes this week to act on it.

And send it to someone who keeps asking which AI they should use.

Share

PS: This newsletter grows from your shares. Sharing is free & keeps these guides free, too.

The one rule

If your company pays for Claude Enterprise, use Claude properly.

If your company pays for ChatGPT Enterprise, use ChatGPT properly.

If your company pays for both, use both.

If you can afford both (and you’re the CEO paying), pay for both.

If you don’t write in English, I can’t help you. I only know English & French. And I wouldn’t trust benchmarks online. The real benchmark is you testing the model.

But Gemini seems to be winning most of the time:

I know you’re not happy to read this. You read me to cut through the noise.

But the market doesn’t care about our need for simplicity.

Models are shipping like crazy. And it won’t stop anytime soon.

The following sections are designed to help you (1) decide and (2) know the how.

1. Claude stays my work brain

Claude is still my default. Especially Claude Cowork.

The reason is boring: folders.

My folders have my files, acting like super-prompts to keep doing the work well and reliably. And I point Claude to this folder on my computer.

Inside that folder, I have files like:

• about-me (who am I, what do I hate and love),

• my-company (what do we do, our mission and goals),

• skills (shared with my team to keep doing the same thing, like /linkedin-caption).

• claude outputs (nothing here, but just claude’s work)

Claude reads the files before answering. It has all of the context needed.

And context is the whole game, so you don’t repeat yourself forever.

Here is the (only) prompt I use:

I want to do [TASK] for [SUCCESS CRITERIA].
But first, read my folder and use AskUserQuestion tool.

AskUserQuestion = a tool where Claude is the one prompting you.

That prompt is simple. It’s also probably better than 100 “magic prompts” you saved and never used again.

And my Claude is connected to my entire stack of tools: Gmail (emails), Slack (team discussion), Notion (where my writing lives), Granola (meeting notes)…

It was all so simple from December to April because Claude was the answer to everything. But starting today, ChatGPT is back.

2. ChatGPT is back.

I didn’t expect to write this sentence so soon.

ChatGPT is back.

And the thing that pulled me back first was images.

Their new image release is better than any other AI:

And if you’re asking, yes, I used ChatGPT to make this image:

How to access the new ChatGPT-image.

1. Go to chatgpt.com.

2. Make sure to turn on “Thinking”.

3. Give it a prompt, and that’s it. You’re using the new ChatGPT:

This is a new kind of model. Definitely a leap.

Here are the best prompts I could find from my own tests & the internet.

Prompt #1: An entire brand guideline.

[UPLOAD IMAGES AND THEN PROMPT] Create a multi slide brand deck for this brand. include social media templates and website mockups

Prompt #2: Make an infographic… in French.

Technical diagram for Gastroenterology, in French. Something worthy to put on a doctor's office.

Worth noting: ChatGPT now designs in many, many languages.

Prompt #3: Posters in the street.

Make a wheatpaste poster setup on a brick wall in sf featuring posters from various AI labs.

Prompt #4: Marketing in every ratio.

Generate a high-tech VR headset exploded view with detailed component labels and promotional copy. Clean high-tech 3D rendering, studio lighting, glowing accents, soft purple-blue gradient background. Vertical stacked exploded view of 9 layers with precise callout labels on left and right sides. Ratio: landscape.

Change the ratio. You can click on any design and change the ratio inside ChatGPT.

Prompt #5: A caricature of you.

[UPLOAD A PHOTO OF YOU] Create a caricature of me and my job based on everything you know about me.

Prompt #6: Create a mockup of Linkedin.

A cinematic 16:9 hyper-realistic shot of a laptop, slightly angled. It's the Linkedin timeline that talks about OpenAI new ChatGPT image model.

Prompt #7: Any recipe, beautiful.

Image of "How to make almond croissant: step-by-step guide".

Prompt #8: Realistic photos.

A 16:9 hyper-realistic shot of a creative agency presentation room. Mood is warmer, more lived-in: sketches on walls, laptops, coffee cups. The main screen shows a designed slide with editorial composition: layered typography, overlapping elements, hierarchy. Text reads: 'we had no idea what to present' 'so we asked AI' 'client loved it' 'we’re still confused'. Natural daylight entering the room, soft magenta shadows. Realistic screen rendering, no distortion.

Prompt #9: Fake iMessage message.

A photo (using an iPhone 7) of an iPhone 7 showing an iMessage conversation between me & my mom scolding me because I played video games last night until late. And I'm very sorry using old-school emojis. 

Prompt #10: A full course on Newton.

create a series of images of college-level textbook pages summarizing and demonstrating newtons math and major science contributions.

Prompt #11: A map of NYC like Ghibli.

NYC Stereotype map by district in Ghibli.

Prompt #12: Nostalgic photo filter.

[UPLOAD A PHOTO] Apply a nostalgic filter on it like I shot this with a disposable camera and a big flash; it's nighttime.

Share

3. ChatGPT wins Excel.

This is the part I think most people will use.

Uploading spreadsheets into AI has always been clunky.

That’s how I used to make Excel with AI:

1. I would go to Claude Cowork. Ask for an XLSX. file.

2. I then open a Google Sheet and upload the one from Claude.

3. And I take it from there, without being able to use AI to modify it.

ChatGPT just released an extension inside Google Sheets. It’s different:

Open a real sheet, and simply open ChatGPT on the side:

And then prompt it on the left side chatbot:

Here’s my favorite prompt template inside Google Sheets:

Create a complete Google Sheets workbook for: [spreadsheet goal].

Context:
[who will use it, what decision/task it supports, and any data/constraints to use]

Build it in the current spreadsheet. Create or rename tabs as needed. Make it usable immediately: clear inputs, linked formulas, summaries, checks, and a clean executive-ready layout.

Rules:
- Use formulas wherever values should update.
- Hardcode only user-provided inputs or clearly labeled assumptions.
- Add a dashboard/summary, input or assumptions area, working model/data area, outputs, and sanity checks when relevant.
- Make reasonable assumptions instead of asking questions unless a missing detail blocks the build.
- Keep labels concise and make the workbook easy to audit.

That workflow is so good because I love Google Sheets: it’s free, collaborative, syncs up automatically, and super easy to use.

Now you can power it with ChatGPT. And the new ChatGPT-5.5 model is on a completely new level when it comes to spreadsheets:

4. ChatGPT is best at search now.

I used Grok for search.

I’m now removing it from the recommended stack for most people (for now).

No one needs another tool to manage.

This is my favorite deep research prompt:

Search the web for the latest information on [TOPIC].

Give me:
1. the current answer
2. what changed recently
3. the best sources
4. what is still uncertain
5. what I should do next

But if you really (really) want to do it with Claude, here’s how:

5. Gemini for non-English work

Gemini still has one job in my stack:

Languages.

I’m in Israel, so I notice the differences in Hebrew (but I don’t speak it well yet).

AI is terrible with small languages because there is not enough text from this language on the internet. The sentence either sounds native or it sounds like someone dragged English through a machine.

If you have to choose the best AI and you work exclusively in your native language, maybe you should consider Gemini. Or the one that you feel is best.

And if you can’t tell → then Claude/ChatGPT, both preferably.

6. Gamma stays for presentations

Gamma stays.

→ I still don’t want to build decks from scratch.

→ But I also don’t start in Gamma.

I start in ChatGPT (best at searching) like this:

7. Use Claude Cowork to write.

Even if the new ChatGPT-5.5 is (supposedly) better than Claude at writing, I prefer Claude. And the reason why is Claude Cowork.

If you had only to read two guides on it, make it these:

8. Where to start.

Don’t test everything. You won’t.

Pick 3.

✦ If you write: → set up Claude Cowork.

✦ If you touch spreadsheets: → install ChatGPT in Google Sheets.

✦ If you make visuals: → test ChatGPT Images.

✦ If you work across languages: → test Gemini.

✦ If you make decks: → test Gamma.

The annoying truth.

You want one AI.

Companies want one invoice.

Teams want one policy.

I understand why. But models are moving too fast.

In December, Claude Opus 4.5 changed my workflow. Then Claude Code. Then Cowork. Then Opus 4.6. So for a few months, the answer was easy:

Claude.

Now ChatGPT 5.5, Images, ChatGPT + Google Sheets, search, and Codex look like the next wave. Maybe OpenAI turns Codex into a Cowork-style for everyone.

And guess who wants a super app too?

Gemini (Google), Copilot (Microsoft), and Grok (xAI, SpaceX).

Maybe Gemini finally turns Google Workspace into the work system it should already have. Maybe xAI does something fantastic with SpaceX, xAI, and Cursor. Maybe Microsoft will finally deliver a good Copilot to its 10 million users.

Models change every 2 months now. I’m not prone to hype. You know that.

But my role is to tell you when a massive switch happens.

And one is happening right now.

PS: This newsletter is free because you guys are sharing it.

On every one of my free articles, I get over 1,000+ shares!!

Share

The best kind of sharing is with your colleagues in your group chat (on Teams or Slack). You’re helping them to adopt AI, and you help me spread the word!

And if it’s the first time you’re here, subscribe to never miss the next:

Subscribe now

They released another Claude:

This one designs.

It lives on another website claude.ai/design. Runs on Opus 4.7 (the best vision model they’ve ever shipped). And it comes with a “Send to Canva“ button:

I have been abusing it since the day it launched. And I understand why Figma (the leading tool for designers) lost $730M valutation right after the news.

Claude Design is another (little) revolution.

So I wrote this guide with 3 goals: (1) the simplest steps to try, (2) the exact prompts that will make you say “wait, I can really design this?”, (3) an advanced workflow I did for myself, that you should copy too.

Two things before we start:

1. Save this guide and spend 30 minutes this weekend to try Claude Design.

2. Send it to any designer, founder, or marketer who still thinks AI can’t design.

Share

PS: This newsletter mostly grows from your shares. And I keep hitting 1,000+ shares! It’s my north star. I now know what you love (or don’t). It’s free & helps me stay laser focused.

1. How to access Claude Design.

Claude Design is a separate product from everything else.

It doesn’t live inside the Claude browser or desktop app.

→ It has its own URL: claude.ai/design.

Cowork still plays a role in my workflow, though. More on that in section 3.

✦ If you’re on Pro or Max:

1. Go to claude.ai/design. Sign in.

2. You do need one of the paid plans.

✦ If you’re on Team or Enterprise:

Claude Design is off by default. Your admin must turn it on:

→ Organization settings → Capabilities → Anthropic Labs → toggle on.

It’s a research preview = so gradual rollout. If claude.ai/design redirects you home, try again in a few days.

And it uses your existing plan limits. Extra usage is available if you burn through the cap. It uses tokens extremely fast. So be cautious.

Share

2. First quick tests (each under 2 min).

My first prompts to quickly test it.

Test 1. Create a stunning website.

I am still in awe of the website I created with… a 2-line prompt.

This is the one-shot result. I didn’t write a single thing on this website.

Here’s how I did it:

1. Go to the first tab, “Wireframe.” Select “High fidelity”.

2. Paste this prompt:

Create a high-fidelity landing page designed to raise $[FUNDING AMOUNT] from [TARGET INVESTORS] for "[PRODUCT NAME]" - [short product description].

Target audience: [Your target audience description]. Tone should feel [how should the web visitor feel] - think a mix of [website] + [website] + [ecosystem].

3. Here’s my prompt for example:

Create a high-fidelity landing page designed to raise $100 million from Israeli VCs for "Governmental Underwater Datacenters" — a sovereign, nation-only underwater data center infrastructure built specifically for AI and robotic workloads.

Target audience: Serious Israeli venture capital firms and government tech investors. Tone should feel premium, strategic, slightly futuristic but highly credible and secure — think a mix of Stripe’s polish + Anduril’s defense-tech seriousness + Israeli startup pragmatism.

Test 2: Let’s make a slide deck now.

I loved my previous idea of underwater data centers.

So let’s make a slide deck now, meant as a pitch deck for sales:

1. This time, select ‘Slide deck’.

   

2. Paste this prompt (if you want my example):

Create a pitch deck for my sales team. Our company has underwater GPU datacenters, meant only for Israeli companies, to ensure full sovereignty.

3. Answer the questions from Claude:

Here’s the result (as a PDF): right here. Impressive for a one-shot.

I still prefer Gamma for its diversity and simplicity. But I’d love a combo of the two. Maybe this is on the future roadmap of the two?

Test 3: Create an animated video.

1. Click on “From template” this time.

2. Write a quick prompt that has a visual story to share. You can also start with a sketch (but I am literally the worst guy at drawing anything).

Here’s my prompt (I used Claude for it, yes):

Create a 45-second animated video explaining how Israel is building underwater datacenters off the Mediterranean coast to cool the next wave of AI compute. 16:9.

Structure:
- 0-5s: Cold open. Title card "The AI compute race is going underwater." Wave animation in the background.
- 5-15s: The problem. Animated stat reveal: "AI datacenters will consume 945 TWh by 2030." Show a row of land-based datacenters lighting up with a thermal overlay as heat builds.
- 15-30s: The solution. Animated cross-section of a sealed server capsule descending off the coast of Tel Aviv. Seawater flows around it as passive cooling. Small fish swim past. Bubbles rise.
- 30-40s: Three stat cards slide in, one after the other: "40% lower cooling costs." "Zero freshwater used." "8x closer to European users."
- 40-45s: Closing card. "Israel's underwater datacenter race. Starts 2026." Leave room for a logo bottom-right.

Visual style: clean tech, thin-line illustrations, Mediterranean blue + sand beige + pure white palette. Kinetic typography for every stat. Smooth camera pans, no hard cuts. No stock footage. No emoji. Serif for titles, sans-serif for body.

3. Get a fully animated video:

Share

3. The video hack (to get great slides).

This one is weird. And it works.

1 - Copy an entire blog from someone (not me, thanks).

For example, here, I use a feature no one else is using: Claude Research.

2 - Paste it into Claude Design.

Upload your markdown file with this prompt:

Make a 30-second animated video that summarizes this blog for a first-time viewer.

You get an actual animated video. Motion, transitions, pacing.

Then, on the same chat, continue with:

Now convert that video into a slide pitch deck.

The slides come out better than if you’d asked for the deck directly.

Because the video step forces visual thinking before you commit to static frames.

Share

3. The advanced workflow: 0 to 1

(Cowork + Claude Design)

Here’s a workflow I run on client projects.

It starts in Cowork. And because Claude Design is only as good as the design system you give it, you need to work on its input first.

Step 1: Extract your brand system with Cowork.

Drop every brand asset you have into one folder. Logos. Past slides. Photography. Landing pages. Your brand PDF, if you have one. Product screenshots.

Open Cowork. Select the folder.

And then paste:

Analyze this folder and produce a full design system write-up. Fonts, colors, graphical styles, component patterns, tone, layout conventions. Flag anything that's missing. Save it as DESIGN.md in my folder.

Cowork reads every file. Outputs a clean DESIGN.md.

DESIGN.md is just a fancy way to say “brand guidelines & instructions”.

Step 2: Upload DESIGN.md into Claude Design.

1. Go to claude.ai/design.

2. Drop your DESIGN.md as context.

3. Every future prompt applies it automatically.

4. You never re-specify colors or fonts again.

(If you already have a design system in code, link the codebase directly. Same effect.)

Step 3: Generate.

Go back to claude.ai/design and upload your DESIGN.md file.

Then add your prompt + your format.

Every good prompt has 4 locked-in inputs: goal, layout, content, constraints.

Build a pricing page for [product]. 3 tiers, annual/monthly toggle, sticky CTA on mobile. Mobile-first responsive. Use our Primary Button component. Match the tone of our existing homepage.

In my tests, when the design system is uploaded and the prompt is specific, Claude Design shines. Otherwise, you’ll see the same design everywhere.

Step 4: How to iterate.

Structural changes → go in the chat:

Show me 3 alternative layouts.

Then you need to click on “Tweaks”.

And it creates multiple versions… in seconds:

Pixel-level changes → go on the canvas:

→ Click the edit button. Select exactly what you want to change.

Before a risky experiment → save a branch:

Save what we have, and try a completely different approach.

Step 5: Validate.

Before you export, run these 3 prompts:

Review this for contrast and accessibility. List any WCAG 2.1 AA violations with exact fixes.

Generate desktop, tablet, and mobile versions.

Suggest 2 A/B test variations of the hero section, each with a different angle.

Step 6: Export.

Click Export (upper right). Options: Send to Canva, PPTX, PDF, standalone HTML, or a bundle for Claude Code.

My favorite should be “Send to Canva”… but it does not work:

So I can go from an empty folder to a shippable design in… I’d say about 1 hour for a single page. 2-3 hours for a full website with multiple tabs. Depends how peculiar you are with details.

Before Claude Design? The same work could take me at least a full day.

And that’s without using my next section favorite hack:

Share

4. Copy any design.

This section blew my mind the most.

There is a free website called getdesign.md:

And as you can already guess… you can literally copy the DESIGN.md file (the brand guidelines for Claude) from most brand you know.

Like, for example, you like the Mastercard branding?

You can download it → Give it to Claude Design → Design like Mastercard

Now let’s try to do it, but like Airbnb:

Or like Ferrari:

Share

5. Taste is all you need.

The future is coming at us, very very fast.

• Claude Design builds 10 dashboards in 10 minutes.

• It applies a design system automatically, suggests 3 layout variations, exports to 5 formats.

• You can upload any style to it → it copies it instantly.

What it cannot do is tell you which of the 10 dashboards to ship for your specific users at this specific moment.

That’s taste.

Taste is the ability to say no to 9 versions and yes to 1.

Taste is the override switch. Taste is looking at a critique and saying, “Claude is wrong about this one, because my audience expects minimalism.” Or the opposite.

That part is still you.

Sure, Claude Design is beautiful.

Sure, Claude Cowork writes so fast & accurately.

Until everyone designs and writes the exact same way.

Think about it: you are now paid to do better. To be ahead of the “default from AI”. That’s why I always say “Master AI before it masters you”.

You must be better than AI’s average, AI’s default. Be ahead.

If you’re a designer reading this and feeling nervous about being replaced, you’re reading the moment wrong. Designers with taste are about to have the best decade of their careers. Tools just got cheap. Taste just got much more expensive.

Share

Claude did not pay me.

And I won’t accept any future sponsorship from AI labs.

I simply can’t.

I must stay as neutral as possible, saying exactly what I want based on my own experience testing the tool. And you know what?

Claude Design is far from being perfect:

• It consumes too many tokens (= money).

• It is still very buggy, left and right.

• You lack control.

But I can see how things are evolving…

1. Claude Code changed how we code.

2. Claude Cowork, a few months later, came at knowledge workers.

3. And now Claude Design is coming at designers.

So no, I’m not paid to make this newsletter (or the next).

But I’m sharing, twice a week, how my work is transforming (very fast) with AI. As I’m trying to keep up, I want you to keep up. So we move just as fast.

I am growing this newsletter purely from your (free) shares.

If this article helped you, be helpful to someone else too:

Share

It does not cost you anything to share. And sharing is caring :)

If someone did send you this, thank them and subscribe for free here:

Subscribe now

Not even once.

This free guide solves this gap. Claude For Dummies.

Skip this if you know Claude and have tried it.

For the “pros”, I added a very last section “How pros use Claude” at the end.

You never tried Claude because you either:

• Don’t really use AI.

• Or you just use ChatGPT.

• Or worst (!!!), you only use the FREE ChatGPT.

This guide is the easiest way to get started with Claude.

I will assume you don’t know how to use AI. And by the end of this (very long) guide, you will know more than 98.75% of the world.

Sounds like a good deal? Cool.

Two things before we start:

1. Save this guide. Block 10 min this week & try Claude for the first time.

2. Send it to anyone who has never tried Claude (& still thinks it’s a guy’s name).

Share

PS: This newsletter grows from your shares. And I keep hitting 1,000+ shares! It’s my weekly north star. Sharing is free & helps me stay laser focused on mastering AI.

1. What is Claude?

Claude is an AI you talk to.

You type something, it types back. It can write, think, summarize, analyze documents, work with your files, and help you make decisions.

It’s made by a company called Anthropic.

If you’ve used ChatGPT before, Claude is the same kind of thing. If you’ve never used any AI, Claude is a great place to start.

How it actually works (in 30 seconds)

You don’t need to understand the engineering.

But 3 concepts will save you from the most common beginner frustrations.

1 - Auto-complete at scale.

Claude predicts the next word, billions of times per response.

That’s how all AI assistants like this work. It’s pattern-matching at a speed that feels like thinking, but it’s not thinking the way you do.

This is why Claude sounds confident even when it’s wrong.

2 - Sycophancy.

Claude is trained to be helpful and agree with you (= sycophancy).

Side effect: if you say something false, Claude might nod along instead of correcting you. Don’t trust agreement. You are the one deciding directions.

3 - Tokens.

Claude reads and writes in chunks called tokens.

A token is roughly a word. Every conversation has a limit on how many tokens fit. This is why long conversations eventually becomes too much: the memory fills up.

★ Remember. Claude is an auto-complete machine at superhuman scale. It sounds sure of itself because it’s made to agree with you. It uses tokens to understand you and memorize your conversation (until it can’t anymore).

Share

2. Claude vs. ChatGPT

Everyone’s heard of ChatGPT, even people who don’t use it.

So let’s take it as the reference point.

Claude and ChatGPT are the same species. You talk to both of them the same way. You can ask both of them the same questions. But they have different personalities and strengths (and trainings).

Here’s what’s actually different:

• Writing voice. Claude’s default writing is less AI-flavored than ChatGPT’s. If you’ve ever noticed ChatGPT sounding like a corporate memo nobody asked for, you’ll notice Claude does it better.

• Long documents. Claude can read a 200-page document in one go without losing track. ChatGPT has been catching up, but Claude still wins here.

• File work. Claude’s desktop app can see your local folder (from your computer) and work with your files directly. ChatGPT can’t (yet 👀).

• Multi-step jobs. Claude has a mode called Cowork (we’ll get there in section 8) that runs tasks for minutes to hours. ChatGPT has agents. But they don’t match Cowork’s scope yet, and they are too technical. Cowork is simple.